This was rolled out RIGHT away, as when I just went to use Gemini, seeing 3.8 Flash be replaced with "4 Hydrogen" was less hype and more a thought along the line of "What [on Earth] is Gemini Hydrogen???", then me going here to figure it out.
This model naming scheme works for separate products, but using elements, molecules, etc. for products you're supposed to compare with each other... Not very helpful.
If anything, my intuition would say that Hydrogen is more powerful than Argon... I don't know. Not liking this naming scheme trend.
OMG, it rolled out on my paid Google Workspace account already too! What a change from ~"wait 8 weeks or more from google.com rollout to <yourdomain> rollout"
You'd think so but they've been updating gemini.google.com really slowly. Up until couple weeks ago it still had gemini 3.1 - which is a February release model.
What do you mean? I have gemini hydrogen flash and thinking and 3.1 pro on my workspace account today. I'll just use another provider with less slowrolling of frontier and better tool usage.
If one of the dimensions it improves upon is inference cost per unit of work, then switching over to it would save money, and the sooner you do that, the more you save.
Yea, it’s Terra 6.1 and Sonnet 5.5 level performance, per the benchmarks. Like you, I was left wondering whether this was better or worse than Argon, etc. It’s not intuitive; gotta read the fine print.
According to Artificial Analysis, Gemini 4 Hydrogen (54) is closer to GPT-6.1 Sol (55) than GPT-6.1 Terra (51). Interestingly, it's also better than Gemini 4 Argon (50), which was Google's previous Pro model. Claude Sonnet 5.5 (55) is slightly ahead though.
Problem is we have no idea here what it should mean.
Carbon is the next version of Argon. But Hydrogen is supposed to coexist with Carbon. Also no idea where the scheme is coming from, what elements should we expect.
Is Carbon the next version of Argon? My understanding was that Argon is the middle-tier model (somewhere between OpenAI's Terra and Sol in position) whereas Carbon is supposed to be like Astra. This is assuming they don't have an even higher tier model upcoming, in which case Argon ~ Terra, Carbon ~ Sol and Element X (Plutonium?) ~ Astra.
Wild guess: maybe they are going with how abundant the elements are to scale the models. I'm no chemist so I don't know reliably how the relative abundance of Argon and Carbon works out, but Hydrogen is "well-known" as the most abundant element in the universe, so maybe that's why its the base model. But, this is essentially a trend of one, so not a trend at all, unless someone can verify it still works for Argon and Carbon.
If you look at Google's API pricing, Gemini 4 Carbon and Gemini 4 Argon are the same while Gemini 4 Hydrogen and Gemini 3.8 Flash are the same (ignoring promotional pricing).
Intuitively, Argon is 4 Pro, Carbon is 4.1 Pro, and Hydrogen is 4.1 Flash (slightly better than Argon according to benchmarks).
Also I intuitively get that a sonnet is "shorter" than an opus (although Fable is less clear) - I have no idea what i'm even supposed to be comparing with google elements... Atomic number? I know hydrogen is the lowest and i guess argon is bigger than carbon but am I expected to know that? Or is it by weight? Argon's a gas so it's lighter than Carbon?
The sol/terra/astra, haiku/sonnet/opus clicked without needing to be explained
> Carbon is the next version of Argon. But Hydrogen is supposed to coexist with Carbon. Also no idea where the scheme is coming from, what elements should we expect.
The problem with using elements names simultaneously as both version numbers and product tiers, is that inevitably you will face a confusing collision or you will exhaust the space.
I don't think it's good scheme to follow but google is just following the trend of anthropic and OpenAI set up. I don't think anyone can understand Anthropic's Model naming scale with Haiku, opus, fable and Mythos. Similar with the OpenAI tera and other models in their line up.
This is so dorky, but I have a comparison with frisbees. In Disc golf each brand does its own naming scheme related to the speed / stability of discs. One company (Prodigy) produced a _roughly_ ordered approach and people very regularly commented on how confusing it was. By comparison, other brands (which got no critiques) have literally arbitrary names matching only a theme (e.g. electricity, animals, ...mental disorders/phenomena). It took me a very long time to come to grips with it. I ultimately think with branding... it only matters that people who use the system can identify things as distinct. I think the subset of minds that wants a strictly logical relationship to ordering is fleetingly small. To most people its additionally not even relevant.
My only real complaint with Google is they regularly change names for seemingly no reason. I'm still out here saying "Gemini" when I type "agy" and I don't bother saying "Anti gravity" because I assume that name won't exist in a year or two.
I think the others are mostly pretty clear. A haiku is a very short poem; a sonnet is a longer poem; an opus is an artistic masterwork ("Magnum opus"), so not necessarily longer but certainly grander. Mythos is slightly more tenuous but still goes in the direction of "cultural awe", something that the highest forms of writing aim for. Fable isn't a perfect name TBH, it sounds much less grand than "mythos" to my ears despite being roughly the same product AFAICT, and with a dangerous hint of "I might make things up".
I think OpenAI's names are clearest. In most people's understanding, moon < Earth < sun < stars.
I think of it as earth being where I am, not in relation to the size of the items, so to me it’s weird that Luna is far away, then Terra is close, then Sol is further. So even those names aren’t clear to everyone. Actually I didn’t realize the ordering was size-based until you said it. And Astra isn’t really a size, is it?
More importantly, Anthropic has been consistent with their naming convention for years (two, yes, but plural years) now. OpenAI and Google’s habit of changing naming conventions semi-regularly makes their products harder to grok.
> but google is just following the trend of anthropic and OpenAI set up.
They botched it.
A haiku is a short poem. A sonnet is a longer work. An opus is a great work. Mythos is a connected body of works.
As with Open AI, Luna --> Terra --> Sol --> Astra from smallest to largest. Very straightforward.
With Google it's totally incoherent. Hydrogen --> Argon --> Carbon!?
Hydrogen is small, but it's the most abundant element in the universe by far. Argon is >3x larger than carbon by molecular mass, though carbon is more abundant. It's impossible to determine what's supposed to be smaller and what's supposed to be larger.
It's funny how different naming schemes "click" (or don't) with different people in different ways. When I first heard the haiku/sonnet/opus it was immediately obvious that they were linking the size of the model to the size of the piece of writing. Same with moon/planet/sun etc. for OpenAI's models.
If Google was going to go with elements, it should have been an absolute no-brainer to link size of of the element to the size of the model eg. Hydrogen/carbon/argon, but no, Google screws up the naming like they _always_ do. Hell, use the noble gases for christ sake. Helium/neon/argon/krypton/xeon would have been great, not perfect mind you, because a lot of people might not remember the ordering of those elements, but at least it would be consistent and make sense.
Yea, impossible to know whether it's supposed to go by atomic mass, frequency or some other metric. Plus most people off the top of their heads don't really have these stats memorized. With OpenAI and Anthropic it's way more straightforward. If you stop the average person on the street and told them to order the three by mass, most couldn't do it. But they could order Moon/Earth/Sun/Astra
Yeah. Atomic mass: Makes no sense, argon is larger than carbon. Frequency: Makes no sense, argon is least frequent of the three and hydrogen is everywhere, so carbon would be in-between rather than the flagship.
Letters of the alphabet: Also makes no sense. There was a theory that Argon = A, that there was an unreleased B model (Bismuth or Barium or something), Carbon = C, etc. (Similar to Intel's MMORPG-inspired Alchemist --> Battlemage --> Celestial...) Hydrogen broke this notion.
Now written out like this, I can see a D = Deuterium, which is mass 2 or heavy hydrogen, but somewhere along along the line it was lost in translation.
The average person isn't supposed to care, they won't be switching models to prompt for the best television for under $1000. But for professionals they should have some sort of understanding of these things.
> I’m not sure if the average person on the street would know the terms Luna, Terra, Sol, and especially Astra.
I don't think these are as difficult as you're making them sound? I think the majority of people who will care what model they're using will know "lunar", "terrestrial", "solar", "astronomy" etc?
Google genuinely have made a huge mess with their AI product stack. They have Antigravity, Antigravity IDE, Gemini app, AI Studio, etc. They have too many similarly named and confusing products. Google should seriously fire their branding team and bring some new talent to consolidate and restructure things logically.
when you put it this way, it sounds like they are operating on an XY axis (at least) vs a continuum. Argon big, complicated, slow AND expensive; Carbon abundant, widely available, can do anything, costly as a result; Hydrogen, abundant, volatile, cheap. can't imagine how insane Nitrogen must be
I find theirs to be a lot more excusable, given it's obvious that the Sun is bigger than the Earth, which is bigger than the Moon. Anthropic less-so, but if you know the meaning of each, you can infer it somewhat well.
What about it makes no sense? I find it to make the most sense, given it's about text generation and their models are named from shortest text to longest.
They aren't, though- a fable is only bigger than a sonnet. Technically, an opus is any creative work, including a music score. A magnum opus would be a person's most celebrated work, but that doesn't have to mean biggest or longest.
I think people on HN argue way too much about this. It's not like they're named iPhone 18 Pro XL Max. It's meant to evoke a feeling of wonder or awe, not really that hard of a logical leap to make, and it's only one minor quibble. Jeesh.
The general trend -- naming models classes after objects from some implicitly ordered set -- was started by Anthropic. It's generally fine, but they didn't leave enough headroom for their largest models, so the implied ordering for opus/fable/mythos is kind of muddled to the average person.
Small/medium/large/etc. has the most easily deducible ordering, but it's not great from a marketing perspective and also has headroom issues unless you're ok with names like gpt-8-xxxxl.
OpenAI's naming is probably the best. The ordering is easily deducible and there's plenty of headroom in both directions. Tons of cool sci-fi sounding names still on the table, and it also comes with the implication that they're approaching a singularity.
I don't know what Google is doing. It's kind of impressive to have an inconsistent ordering after only three names.
Same, my eyebrows about jumped off my face. To go from Google as normal -- product announced, rolls out randomly in some access modalities, mysteriously doesn't in others because no org talks to any other org -- to "wait wtf is this new model I didn't even hear about this" is so laughably bizarre. Not going to get used to it, I know Google.
Agreed on the naming issue and while it was novel at first it will inevitably get worse as more models get release, until decide to abandon it all together. It’s like they are setting themselves up for failure on that front lol
Look at how many AI adjacent products Google talks about in this press release along with its model names. Anthropic and OpenAI are guilty of it as well, so many models and ways of working with them right now, letting the users figure out the nuances of each.
Earlier today I found myself asking ChatGPT to research and explain exactly what Anthropic Playground is for and how specifically it may differ for my tasks from Anthropic Workspace, Chat, Code, and API credits. I chose ChatGPT Sol 6.1 high for this task going on gut feel, hoping it was the best model to handle that job without wasting my precious Claude tokens.
I think there need to build these adjacent things to create lock in. I can switch between Claude, GPT and pretty much and model on OpenRouter in little time, especially for dev work. With all of these being weeks apart in progress on the models and open-weight models seemingly getting closer and closer there is almost no moat. Without a moat they cannot charge high prices - it's a commodity. Having lots of my data in their system and having other exclusinve features like Claude Design or Whiteboard and what not, they create something that I cannot keep if I switch to a different vendor for a cheaper model.
Yes the naming is nonsensical, because the intelligence ranking is this (with atomic numbers in parentheses for those that don't have the periodic table memorized):
Carbon (6) > Hydrogen (1) > Argon (18).
Double nonsense because it appears Hydrogen is in fact a smaller model than Argon, in which case this should obviously be released as "Gemini 4.1 Hydrogen" or something.
What they probably should've done if they really wanted to use elements is to do noble gases, and start these first three models in the middle (to leave room for bigger/smaller models in the future).
So you'd have:
Argon => Argon (18)
Hydrogen => Krypton (36)
Carbon => Xenon (54)
Not exactly because as noted it appears hydrogen is in fact the smaller model, but then this should be (again) 4.1 or something.
And they'd leave space for a truly massive model if needed (Radon) and two slots for much smaller models, with Helium (2) presumably being a model that could be run on basically any consumer device.
As this is a) being served in Search, b) supposed to be 2.5x faster than 3.8 Flash, c) the element with lowest atomic weight; I deduce that, despite being more capable than 3.8 Flash, now the “entry level” model set to replace 3.5 Flash Lite.
But it is priced the same as the non-discounted 3.8 Flash, considerably higher than what anyone is paying today for 3.8 Flash or 3.5 Flash Lite or the deprecated 3.0/2.x Gemini models. How does this work? Is Google abandoning the “fast and cheap” tier? That seems like a mistake since plenty of workflows are absolutely price inelastic, meaning when the 3.8 Flash discount ends at EOY, lots of current enterprise Gemini customers would have to take their business to another shop??
Same impression here. If this is Google's strategy we'll have to migrate to Luna, because costs are already high for our workload on 3.8 Flash (which in itself was already a 40% increase compared to 3.7 flash due to being more verbose and thinking more) and this now doubles cost once again.
Depends on your needs. Gemini Flash has best bang for your buck on visual tasks and best multilingual support/intelligence for non-Western/non-East-Asian languages; better than paid or OSS alternatives.
Yes I gave for granted textual tasks but they're not the only ones. If I remember correctly, mistral large should be good as well on image tasks while being cheap(er). I never had to do anything with less common languages tho
According to Artificial Analysis, it's much cheaper per task than 3.8 flash, 3$ for H xhigh vs 6-7$ for 3.8 medium-high. I don't understand how though, since AA says it has more output tokens per task.
This is based on what we measure from the API - we see different endpoints depending on how they are served can have different cache hit rates (cache hit rate is defined as the % of cacheable tokens that actually come back as cache hits)
Anyone else got suspended for just using agy on the command line? "This service has been disabled in this account for violation of Terms of Service. Please submit an appeal to continue using this product.".
The support claims that it can do nothing, not even see a reason and pointed me toward a form with no text inputs just a checkbox along the lines of "I understand the TOS, please check". So I filed an appeal, and it seems that it was automatically processed and within a few minutes. agy was working for one more prompt and it triggered a suspension again.
What channel can I use to get any attention on something like this? I've heard these Google horror stories... fortunately this is only the Antigravity part, but I've just today upgraded to Ultra and this coincided with that. I've become extremely hesitant to continue using broader Google services when this can not be resolved via their customer support.
I am also spooked. I reached out to support once more, and they seem to have escalated this to backend support (seemingly working on a fix live?). So this seems to be some larger issue and not a "typical" false positive. If I get to know anything about the reason, I'll follow up here.
The thing that preceded this was an upgrade to ultra today from pro, and a call of agy cli from within opencode (explicitly not using it via some login/oauth, but just calling `agp -p "prompt"`, which had been working for months). Right now it seems that every agy chat works for 1-3 messages, then gives a suspension error.
If this gets resolved via live/email support timely I would be fine. But I am spooked by the appeal form/process that is completely opaque via a google form(!) and seemingly can can take days: https://docs.google.com/forms/d/e/1FAIpQLScOJWibQ_-hYuVv63kJ...
I used agy for apps and had no problems... However, the idea of this happening never occurred to me. It might make more sense to have my AI services under a different account :-(
I want to go all in the Gemini ecosystem but several things keep me going back to OpenAI:
* the Gemini app is glitchy on iPhone (my chats frequently fail or are lost if I background the app)
* the Work equivalent (I guess task?) is not as powerful. I often have ChatGPT Work to arbitrary computation and even install ML on the short-lived VM.
* there is no simple integration between antigravity and the Gemini app. I want a way to continue working away from my server while using the familiar app. No I don’t want VNC.
* ChatGPT sites are cool for putting up proof of concepts with zero effort
I believe this has been rectified and now if you get caught breaking Gemini ToS, it just bans you from Gemini (I think you might even get a cool-down warning ban first).
For more egregious violations like the kid uploading naked pictures of himself, I don't know.
Glad to see this. However, given the general level of product confusion in the Gemini lineup, I'm not planning on relying on this even if it is their intent.
> there is no simple integration between antigravity and the Gemini app. I want a way to continue working away from my server while using the familiar app.
On the other hand, I'm very glad they have separate quotas.
FYI, you can use Antigravity as a web app with "remote control" mode. It works pretty well, better than VNC since the page is responsive, though the text is pretty small and the UX if not really differen than the computer version.
I've found agy remote control doesn't work anywhere as nicely as claude code. I tried using agy on my PC, enabled remote control, then opened on my Android browser. If I type anything, there is no way for me to delete (backspace) if I make any typos other than refreshing the page. It basically made the experience unusable for me. I don't understand how something so basic as the backspace button can be broken. Every time I try a Google AI product something very basic is broken. PDF uploads on the Gemini app also don't work nearly as well as with ChatGPT or Claude. These are all things I'm sure Gemini Carbon can find and fix in 1 day. I don't understand how they ignore basic UX to such an extreme.
If I had to guess the backspace bug is usually the Android keyboard. Gboard sends composition events instead of keydown events, and a lot of web UIs only listen for keydown. Switching off predictive text or trying a different keyboard fixed it for me.
Otherwise you can use agy on the Cosyra mobile app, its far better. I'm not going to claim its better than the Claude app, but its certainly better than the Gemini RC.
It will be the same as these smaller models from OAI and Ant. Strong in benchmarks but pretty bad in real world use.
I've read this same comment every time a new Luna came out, and the difference between free GPT (Luna) and paid (Sol and Astra) is still outrageous for general use.
Maybe I’m spoiled but I found it repulsively dumb. Some of the highest rates of hallucinations since GPT-3. Incredibly lazy. Heavily biased to parametric knowledge and speed over accuracy. I would rather have no AI at all than use that.
It's definitely not the case that small models only do well in benchmarks. We're working on a battle system on top of OpenCode, so have data for real people doign real work.
People keep saying this, but inference itself is not what is being subsidized. The expense for electricity/hardware/datacenter operations/etc for an inference token is less than what they charge for it, so that cost doesn’t HAVE to go up to be profitable.
The thing that is being subsidized is the TRAINING costs, and that is because they aren’t selling enough tokens yet to pay for the training costs. This is just like most other growth businesses, though, and if enough people use enough inference, it can be self sustaining.
Price per token is dropping (though anthropic are doing dastardly things such as using new tokenizers that use 30% more tokens), efficiency will likely increase.
If the subscriptions (of all of the big companies - unlikely?) give a poor amount of value in the future, will move to trusted/open router and use a harness that uses the latest cheap good models for most things, more expensive models as orchestrators.
Already there are non-SOTA models that should be sufficient for most non-coding tasks, and soon for coding, also.
The future is bright, in that respect, without subsidizing the ridiculous IPOs.
I understand his assertion, and I can make up an argument that fits it too, probably a few arguments that aren't even compatible with each other, but what you said isn't what he said, you are just assuming all of this, and also nowhere here does it explain how this fits an analogy of "post nut clarity".
Why is this exchange reminding me of that You Can See Everything documentary?
The analogy made sense to me: Tokenmaxxing policies were the result of an irrational drive by execs who just craved an, ahem, explosion of metrics due to the excitement over new technology. Now that cost sensitivities are setting in, there’s a growing sense of unease that maybe they were in an altered state when they created the mess and now that the excitement has passed they have to clean it up.
So people had this drive to use tokens, and once they experienced the release from using the tokens they went into a refractory period where their reasoning is no longer being overridden by the need to use tokens? And later I guess they'll start to feel more and more of a need to use tokens and this will repeat?
Or is it just a throwaway phrase that makes no sense?
We’ll see if the cycle holds true for tokens or LLMs specifically, but for sure there is a larger “new sexy tech” cycle with a strongly emotional output===outcomes metrics irrationality phase to it.
I think he's just saying tokenmaxxing is a temporary hype lol, probably fueled by subsidized subscriptions. When that's over, these cheaper models might get more usage.
Can anyone understand the logic behind the names? Argon, hydrogen, carbon, what is the mental model one should use to understand where they fit, or what comes before/after?
It's not, Gemini 4 Carbon is better than Gemini 4 Argon.
Argon internally was called Barium (Argon was a different model) so the models were originally named in alphabetical order. Then they change their mind and decided to call the lightweight model following the atomic weight so Hydrogen after they have released (A)rgon and (C)arbon. I imagine that this lightweight model is the original Argon.
My guess, by atomic weight. Hydrogen = "sonnet". Carbon = "opus". Argon = "Fable" only argon 4 just happens to be kinda shit? But also it's google so who the hell knows. They could be just going alphabetical argon, then carbon, then hydrogen as release order follows the alphabet. That'd be kinda dumb but so is their access/subscription strategy so it'd be on brand.
Personally I found my usage of llms bimodal, I only have uses for the top frontier models (for research and development) or for the cheapest and fastest ones (to use in production for consolidated, simple flows). I am curious what people find the mid tier models useful for.
I feel like people use medium models thinking they are saving money, but in the end, it ends up wasting more time and tokens theough the increased rate of issues/bugs.
If the highest tier model makes even one or two fewer mistakes that require human intervantion, that almost makes it a no-brainer choice for production.
I've been doing some research lately on my own (aside from the major benchmarking going on) and I've definitely found cases where medium models do much poorly and cost more than the higher models. But on the other hand, if I can give the medium models certain skills or other resources to complete a task then they perform better in cost than the higher model (which, given the same resources, sometimes gets more confused than the medium does).
At least for my use case of evaluations of a set of features.
You’re talking about something else (reasoning levels). The parent comment is referring to e.g. GPT Terra as opposed to Luna, Sol or Astra, Claude Sonnet as opposed to Haiku, Opus or Fable, or Gemini Flash/Hydrogen as opposed to Flash-Lite, Argon or Carbon.
Oh, good catch. I do use Sol in production, for cases where I need quite good results, but can't afford to run in production the highest tier model due to pricing (so it's like the highest tier model I can afford for that use-case).
To be more specific, this is for a support automation, where the model has to be quite accurate in following rules and generate responses, and Luna models or lower-tier models often make bad mistakes and generate wrong responses, so are not really usable for sending data directly to the user.
Even better at the local LLM stage, where QFN can catch a mistake in prior work and patch it without any subscription costs at all. then I can use a cloud model for polish and finish and move on.
I'm building stuff at home with my work focused on managing infra. If anyone wants to sponsor a SOTA model subscription I will check it out. Until then I'll see how flexible a Strix Halo setup and my homelab can be.
Asking people to guess beforehand how much intelligence they require for a task will hopefully be a relic soon. We’re lacking useful routing techniques which accurately pick the right model for the job. Or a high intelligence model which completes simple tasks fast and cheap. Or something else.
I have some tasks that deserve a bit more reasoning but for a long time most of my needs have been met by mid-range models, rather than the big ones. There's barely any motivation for me these days to use something more advanced than Terra on low/medium for everyday work. For a personal agent doing occasional coding, I'm still on Mimo flash.
I think many people still expect their tasks to need something more fancy, but could do with a reset of expectations and try to go much lower.
Production pipelines for non-coding usage. Jagged edge is very real. Just one example, loads of tasks where Gemma 4 26ba4b wipes the floor with Deepseek Flash v4.1. On Artificial Analysis-like benchmarks, the latter is many times more "intelligent". Not everything in the world is coding and math, but those are easy to "measure". Things that are hard to measure aren't part of their benchmarks.
All the labs are holding back releases now, not necessarily to play games, but to make sure alignment is right. These models are teetering on the edge of "One person can seriously damage digital infrastructure" and an event like "Teenage kid uses Claude to bring Wells Fargo to it's knees for 6 hours" would be absolutely catastrophic for the whole industry.
> Teenage kid uses Claude to bring Wells Fargo to it's knees for 6 hours" would be absolutely catastrophic for the whole industry.
It would (rightfully) be catastrophic for wells fargo. And great for the industry.
If your security model relies on children not having access to things state actors can leverage, your in for a rude awakening.
Right now it's looking like the security model for everyone is "Does or doesn't have access to SOTA LLMs"
Hence why they are gating model releases. Even Chinese models are doing this as Kimi K3 shipped with a lobotomized cyber sec ability. The major labs all have invite only cyber models too.
And hopefully this time for them they don't have to release a new rev the week after.
I wonder now that Terra-6 is out how it would compare to Sol-6. (Which was replaced a week latter by sol-6.1 after chat got dropped opus 5.5 to incredible fanfare then sonnet 5.5 that radically upped the medium model ability).
I have a stupid benchmark: Bongard Problem 2 from here (https://matheminutes.blogspot.com/2019/04/bongard-eggs.html). It's an open-ended visual task that is easy enough for humans, but models are still somewhat bad at visuals, so it's one of the few remaining reliable ways to trigger hallucinations. This one answered correctly, but then felt the need to add a fun fact that was wrong:
> The rule separating the two baskets is layering (which color is on top). Left basket: In every pattern on all six eggs, the blue shape is on top of (overlapping) the green shape. Right basket: In every pattern on all six eggs, the green shape is on top of (overlapping) the blue shape. [true solution]
> (Note: Both baskets contain the exact same distribution of motif counts per egg—two eggs with 2 pairs, three eggs with 3 pairs, and one egg with 4 pairs—serving as a classic Bongard distractor to hide the occlusion rule.) [factually incorrect and also would make no sense as a distractor]
And then trying a harder puzzle just makes it bullshit completely.
I don't understand why Gemini models always add these "it's typical XYZ" bits, which are typically so much worse than everything else in the answer.
The sample problem I kind of gave up on and guessed sort of... it couldn't be sorted by color that's too obvious.
And then the real problem number 1. The answer is "Left - each figure contains an even number of arrow heads. Right - each figure contains an odd number of arrow heads"... but on the right I can clearly see that the three eggs across the top have an even number of arrowheads.
It's not the number of arrowheads on the egg; it's the number of arrowheads on each of the little figures on the egg. Those eggs you mention have 2 figures each, true, but each figure has just 1 arrowhead each.
If you have an Ultra subscription (I don't or I would try it) I would be really curious if Carbon is able to solve your harder Bongard problem without making stuff up given it's absurdly high AA Omniscience non-hallucination score, it's supposed to be like 30%+ better than nearest non-Google competitor. Although Hydrogen xhigh apparently scored within 5% of Carbon so who knows ¯\_(ツ)_/¯
I won't share the solution (you have to keep some mystery in the world), but I showed this a year ago at a high-school reunion and five people independently found the same solution, so it should be well-posed.
Ha, okay I caved and spent the 5¢ to ask Carbon the problem through openrouter and it was indeed able to solve it. No random hallucinated fact either. Thought you might want to know. Cool problem!
Good they are able to replace Flash. And cool this is available to everyone via search. But I still don’t understand Googles strategy with these new models.
For pro sub, Argon is still very limited and Carbon is only available on Ultra. This is not competitive at all. To say nothing of also being locked into Antigravity.
IMO it sits between sonnet and opus in terms of quality, and is equivalent with gpt sol (for this particular benchmark). Google did a great job with this.
How do you actually use hydrogen in the app? The only model selection options I see are "auto," "code and analysis Argon 4" which is paid, and 3 thinking levels.
I really hope hydrogen keeps the excellent conversational ability of 3.8 flash. For brainstorming or even decoding the dense output of the powerhouse models, flash was a real delight to use. Its the first non-SOTA model that I would actually go to often over the larger models, it was fast, plenty smart enough, and generally wrote really well. Here's to hoping hydrogen keeps that up
In the Gemini app of over the API? In the app Ive found it to write poorly, probably a system prompt thing. I've done some stuff with Argon over the API and it wrote well.
This is a good blog post because of the graphs comparing directly to Terra and Sonnet for both wall-clock time and cost per task. This is the exact information I'm looking for.
My TL;DR is: seems like a better choice than Sonnet in areas outside of coding, and at high reasoning it's smarter than Terra but somewhat more expensive.
I'm really enjoying all these Gemini releases. I know a lot of people keep saying Google is too far behind in the race and that they dropped the ball to early but I don't feel that way at all. I had both Opus and Fable try to find a discreet bug in my codebase and both models assured me I had nothing to worry about and was probably overthinking. Argon was the only model that got to the bottom of the matter and developed the fix for me. I constantly find the Gemini models perform better in the real world as compared to the other frontier models and I also find they have less of the AI-isms that many of us have now come to associate with ChatGPT and Claude. Gemini feels like you're speaking to a comrade. Claude feels like you're speaking to a friend and it's an important distinction I feel.
Qwen 4 Pro preview has been beating fable-5.2 in my tasks. This model is available to try on the free tier. I can't even use argon or carbon on my $10 google sub. Why doesn't google show off their top models with a few free prompts a day to get us hooked? Let the dog see the rabbit.
Google are doing such a terrible job. They aren't allowing Pro accounts access to this in antigravity that they gave to students or via mobile plans elsewhere, because they messed up badly and don't have the compute.
They have been absolutely horrible about communicating this and even gaslighting the users, completely nerfing the same models that they do get to use, without telling them that they are doing.
So after being caught up in this, it's left a very bad taste and I'm subscribing elsewhere.
The naming of these Gemini 4 models is ridiculous.
In order of sophistication and cost, the models are Hydrogen -> Argon -> Carbon.
What kind of ordering is that?!?
With the introduction of Hydrogen, it's no longer ordered alphabetically. That would be Argon → Carbon → Hydrogen.
And it never was ordered by atomic mass, which is arguably the best ordering if you're going to name your model classes after chemical elements. That would be Hydrogen → Carbon → Argon.
It's also not ordered by prevalence in the universe. That ordering would also be Hydrogen → Carbon → Argon.
It's not ordered by electronegativity, boiling point, melting point, or any other remotely defensible chemical property I can think of.
It's not even ordered by letter count, Scrabble score, or NFPA 704 Flammability Rating, all of which would be Argon → Carbon → Hydrogen.
Per Business Insider[0], Argon and Carbon were part of an alphabetical lineup:
> An internal document shows that Google has tested a series of Gemini 4 models, internally named Argon, Barium, and Carbon.
Then I suppose they tacked on Hydrogen since it's supposed to be a "light" model and hydrogen is the lightest element. I agree that it's super confusing though.
+1 to this the naming is totally opaque and ridiculous.
There is no hint of what each one does. I get that these were internal code names, fine, but to everyone outside Google it's not helpful to have these meaningless names.
ChatGPT and Claude aren't much better - there is still quite a lot of assumed knowledge there assuming you know what e.g. a sonnet is and what an opus is - but at least if you squint there is something there.
Google's names just seem to be randomly generated.
Maybe they were worried that the 3.1 Pro model being around for so lunch ruined the brand? But 3.8 Flash was pretty decent I thought so it's not like "flash" was burnt. I dunno.
The choice is really odd. Hydrogen and Carbon could have an organic chemistry theme going if they chose Oxygen, but then they go with Argon? Anthropic and Open AI have the names figured out in terms of obvious size: Luna->Terra->Sol->Astra and Haiku->Sonnet->Opus->Fable->Mythos makes intuitive sense. I have no idea how the Google model size class works.
I know some people are still a bit confused by astra, sol, terra, and luna, but at least you can explain the reasoning and it makes sense. Google's new naming scheme is just pure luna...cy.
Here's a comparison grid: https://static.simonwillison.net/static/2026/gemini-4-pelica... - note that Carbon and Argon are priced the same, so it's actually quite neat to see the difference in token usage between them (Google say that Argon is more token efficient.)
Their naming scheme seems a bit tough to follow. But it does seem like they finally learned their lesson and announced something at the same time it's actually available!
Argon, Carbon, now Hydrogen. If Google keeps climbing the periodic table, the roadmap writes itself: Californium finds gold in your codebase (enterprise only)
Claude is the same really. Fable is a short story, so makes no sense in this hierarchy, and Opus have to think it as derived from "Magnum Opus" to make sense, else Opus by itself simply is a piece of music by a composer and would've been more appropriate for low-reasoning model to be used by sub-agents.
Anthropic, OpenAI and literally everyone else's naming schemes are in the same class of silly. The regurgitated complaints on every single thread are really getting tired.
I get that hydrogen is the smallest element but like it's not exactly that intuitive for everyday people to know which to choose. Whatever happened to big/medium/small
looking as the price, seems like google wants to share the AI bowl with other companies, so generous!
though I suspect they keep the price this high to increase the profit margin for the shortened release cycle (instead of 6 months I think they will release newer checkpoints in 3 months or less from now on).
still, my application now permanently use deepseek models instead. no point using gemini, neglectable real life improvements for extremely higher price.
Pretty big Gemini fan, but my initial results with Hydrogen and Carbon are disappointing.
They are trading improvements on pass@1 pointless benchmarks for worse disasters at p90/99 that you're going to hit every week in agentic workflows (if not daily):
overreach (doing more than asked), overwork (when it should have failed fast), harm (doing the opposite of what was asked), orthogonal work (doing something completely unrelated to what was asked), etc.
Google doesn't seem to be the only lab making this trade...
The set of benchmarks they released are disappointing. 10 mostly random benchmarks. Really?
No one on HN seems to get that agentic workflows are not the only name in the game. Everyone doing agentic development is either on a plan or with a custom enterprise offering, either way using higher-performing models. These entry level models are (also) for building apps on top of while you are paying API prices.
Terrible naming convention on the low-end, because what do you do if you make something smaller than Hydrogen next? Gemini 5 Electron?
On the high-end, I feel like there is an implicit promise that if they keep releasing larger models and go past Oganesson (118), they are on the hook for AI-discovering a new stable element, so that is fair at least.
>For everyone: Start using it in the Gemini app model drop down
and then it says for free users there is only "automatic" and you will get Gemini 4 Hydrogen "when it's the appropriate model for the prompt". Yeah sure Google, whatever, I don't care about Gemini models anymore. Too hard to access, constant rugpulls, intransparent pricing, many, MANY outages and in general very unreliable. At this point I just want Google to fail.
Are we sure the guy that creates google naming schemes is not held hostage and those are his desperate attempts to communicate with the outside world in concealed way?
OpenAI is at least arranged by size of celestial feature. Anthropic did arrange them by size of a written work. But I think Fable and Mythos made it a little weird. As for Google, no idea other than Hydrogen is small so at least that kind of tracks.
Having said this, I don't defend this trend at all and also think a model picker ought to be good enough with at most three tiers rather than five or so with sometimes like five effort levels each for 25 combos. I refuse to believe we need this granularity.
Is this a replacement for Flash-Lite? Because it’s being used to serve answers for Search.
The new naming scheme seems shockingly poor though, “Pro” and “Flash” made intuitive sense to me, and Anthropic and OpenAI kept a sense of scale. Nobody is going to intuitively understand “Oh yes, Hydrogen is smaller than Carbon”, even though it is.
Are we going to talk about naming and confusing ways to access it, or are we going to discuss how capable the model is?
Every single time a Google model comes out, an army of Anthropic shills sidetrack the discussion.
Can we look at what the model does and discuss whether Google is still in the game or not? (naming models and whether you need to access it by clicking a drop-down menu or making a prayer ain't really relevant)
Another day, another nonsensically named google product.
I quit using Gemini models months ago because they are incomprehensible to understand in terms of when to use them and how to navigate the terrible GCP and Gemini UI. It’s a Kafkaesque nightmare.
Maybe I'm not on the latest models (I do actually pay the min plan), but just this week I tried Gemini again. It still doesn't even get the current date/time of day correct. If it cant do that, I cant trust any output from it. I have no idea how Google fumbled so badly, but benchmark scores don't mean a thing if you cant even get your LLM to output the current date and time correctly.
You could vibe code a better LLM with Fable than what they have delivered. It is embarassing.
I asked about the time of market open for premarket trading for nasdaq trades. While it correctly determined when the pre markets would open, it then said it WAS open already by referencing the current time which was off by an hour. I told it where I was located and it then shifted by two hours. finally I told it to look it up and it got it correct.
"for all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro and Ultra subscribers in Google Antigravity."
So which one of these six options is the one I should use?
Same question as whether you should be using the latest Claude Opus model from Claude Chat, Claude App, Claude Code (via Claude Pro/Max/Ultra subscription), Claude Code (via API), Claude Platform, Claude on AWS Bedrock, Claude on Azure Foundry, Claude Managed Agents, Claude Cowork, etc?
When you call it through OpenRouter you're using the API.
You know how Claude Code is like a little program that your computer can run which lets you talk to Claude? Well Gemini is actually really similar! If that's more your speed, you should use Antigravity.
But if all of this is feeling really overwhelming and you just want something simple, I would recommend just Googling it! Hydrogen will be the one generating the response :)
This was rolled out RIGHT away, as when I just went to use Gemini, seeing 3.8 Flash be replaced with "4 Hydrogen" was less hype and more a thought along the line of "What [on Earth] is Gemini Hydrogen???", then me going here to figure it out.
This model naming scheme works for separate products, but using elements, molecules, etc. for products you're supposed to compare with each other... Not very helpful.
If anything, my intuition would say that Hydrogen is more powerful than Argon... I don't know. Not liking this naming scheme trend.
OMG, it rolled out on my paid Google Workspace account already too! What a change from ~"wait 8 weeks or more from google.com rollout to <yourdomain> rollout"
Thanks any Googlers who had a hand in this
I'm going to assume they are saving on inference by rolling it out faster.
You'd think so but they've been updating gemini.google.com really slowly. Up until couple weeks ago it still had gemini 3.1 - which is a February release model.
So this is an actual change of pace
What do you mean? I have gemini hydrogen flash and thinking and 3.1 pro on my workspace account today. I'll just use another provider with less slowrolling of frontier and better tool usage.
Why do you think rolling out faster -> more inference savings?
If one of the dimensions it improves upon is inference cost per unit of work, then switching over to it would save money, and the sooner you do that, the more you save.
Here's an explaination of what's going on in the backend.
Alibaba is forced to do this stuff: https://www.tomshardware.com/tech-industry/semiconductors/al...
Meanwhile Deepseek: https://news.ycombinator.com/item?id=45643163#45652636
Smaller model, likely.
If it's more efficient than 3.8 that it replaces.
It's the equivalent of Gemini Flash, judging by the API price. I just tried it in the Gemini app, it's much better than the previous version.
Yea, it’s Terra 6.1 and Sonnet 5.5 level performance, per the benchmarks. Like you, I was left wondering whether this was better or worse than Argon, etc. It’s not intuitive; gotta read the fine print.
According to Artificial Analysis, Gemini 4 Hydrogen (54) is closer to GPT-6.1 Sol (55) than GPT-6.1 Terra (51). Interestingly, it's also better than Gemini 4 Argon (50), which was Google's previous Pro model. Claude Sonnet 5.5 (55) is slightly ahead though.
They just trying to register the model names into our mind. Now a days we remember Fable, Opus, Sonnet. I think Google trying to do the same.
Problem is we have no idea here what it should mean.
Carbon is the next version of Argon. But Hydrogen is supposed to coexist with Carbon. Also no idea where the scheme is coming from, what elements should we expect.
Is Carbon the next version of Argon? My understanding was that Argon is the middle-tier model (somewhere between OpenAI's Terra and Sol in position) whereas Carbon is supposed to be like Astra. This is assuming they don't have an even higher tier model upcoming, in which case Argon ~ Terra, Carbon ~ Sol and Element X (Plutonium?) ~ Astra.
Wild guess: maybe they are going with how abundant the elements are to scale the models. I'm no chemist so I don't know reliably how the relative abundance of Argon and Carbon works out, but Hydrogen is "well-known" as the most abundant element in the universe, so maybe that's why its the base model. But, this is essentially a trend of one, so not a trend at all, unless someone can verify it still works for Argon and Carbon.
If you look at Google's API pricing, Gemini 4 Carbon and Gemini 4 Argon are the same while Gemini 4 Hydrogen and Gemini 3.8 Flash are the same (ignoring promotional pricing).
Intuitively, Argon is 4 Pro, Carbon is 4.1 Pro, and Hydrogen is 4.1 Flash (slightly better than Argon according to benchmarks).
Also I intuitively get that a sonnet is "shorter" than an opus (although Fable is less clear) - I have no idea what i'm even supposed to be comparing with google elements... Atomic number? I know hydrogen is the lowest and i guess argon is bigger than carbon but am I expected to know that? Or is it by weight? Argon's a gas so it's lighter than Carbon?
The sol/terra/astra, haiku/sonnet/opus clicked without needing to be explained
A mythos is longer than an opus.
A fable is a tame mythos.
> Carbon is the next version of Argon. But Hydrogen is supposed to coexist with Carbon. Also no idea where the scheme is coming from, what elements should we expect.
The problem with using elements names simultaneously as both version numbers and product tiers, is that inevitably you will face a confusing collision or you will exhaust the space.
Openai/anthropic's model names have the same issue too
I don't think it's good scheme to follow but google is just following the trend of anthropic and OpenAI set up. I don't think anyone can understand Anthropic's Model naming scale with Haiku, opus, fable and Mythos. Similar with the OpenAI tera and other models in their line up.
This is so dorky, but I have a comparison with frisbees. In Disc golf each brand does its own naming scheme related to the speed / stability of discs. One company (Prodigy) produced a _roughly_ ordered approach and people very regularly commented on how confusing it was. By comparison, other brands (which got no critiques) have literally arbitrary names matching only a theme (e.g. electricity, animals, ...mental disorders/phenomena). It took me a very long time to come to grips with it. I ultimately think with branding... it only matters that people who use the system can identify things as distinct. I think the subset of minds that wants a strictly logical relationship to ordering is fleetingly small. To most people its additionally not even relevant.
My only real complaint with Google is they regularly change names for seemingly no reason. I'm still out here saying "Gemini" when I type "agy" and I don't bother saying "Anti gravity" because I assume that name won't exist in a year or two.
I think the others are mostly pretty clear. A haiku is a very short poem; a sonnet is a longer poem; an opus is an artistic masterwork ("Magnum opus"), so not necessarily longer but certainly grander. Mythos is slightly more tenuous but still goes in the direction of "cultural awe", something that the highest forms of writing aim for. Fable isn't a perfect name TBH, it sounds much less grand than "mythos" to my ears despite being roughly the same product AFAICT, and with a dangerous hint of "I might make things up".
I think OpenAI's names are clearest. In most people's understanding, moon < Earth < sun < stars.
I think of it as earth being where I am, not in relation to the size of the items, so to me it’s weird that Luna is far away, then Terra is close, then Sol is further. So even those names aren’t clear to everyone. Actually I didn’t realize the ordering was size-based until you said it. And Astra isn’t really a size, is it?
More importantly, Anthropic has been consistent with their naming convention for years (two, yes, but plural years) now. OpenAI and Google’s habit of changing naming conventions semi-regularly makes their products harder to grok.
Using the word "grok" here offers some great "Whose on first?" opportunities.
Android version names were random desserts for the longest time. People are expecting too much from Google here
They weren't random. They were alphabetical. The first letter of the name always shifted to the next letter in the alphabet.
• Cupcake (Android 1.5)
• Donut (Android 1.6)
• Eclair (Android 2.0–2.1)
• Froyo (Android 2.2)
• Gingerbread (Android 2.3)
• Honeycomb (Android 3.0–3.2)
• Ice Cream Sandwich (Android 4.0)
• Jelly Bean (Android 4.1–4.3)
• KitKat (Android 4.4)
• Lollipop (Android 5.0–5.1)
• Marshmallow (Android 6.0)
• Nougat (Android 7.0–7.1)
• Oreo (Android 8.0–8.1)
• Pie (Android 9)
They were alphabetically ordered random desserts. It wasn't too bad.
This isn't the same situation. Android versions always had 2 parts: the number and the name. They were always easy to follow.
Using constantly changing names for tiers is a bad idea. Unless the tiers are pre-existing (bronze, silver, gold).
> Fable isn't a perfect name TBH
Fables are short.
They should have used "saga" / "epic" / "cycle" / "series" / "universe".
I realize that you're making a point about ordering here but realistically none of these except maybe Saga work as Anthropic product names.
Epic and Universe for sure work as LLM model names.
> but google is just following the trend of anthropic and OpenAI set up.
They botched it.
A haiku is a short poem. A sonnet is a longer work. An opus is a great work. Mythos is a connected body of works.
As with Open AI, Luna --> Terra --> Sol --> Astra from smallest to largest. Very straightforward.
With Google it's totally incoherent. Hydrogen --> Argon --> Carbon!?
Hydrogen is small, but it's the most abundant element in the universe by far. Argon is >3x larger than carbon by molecular mass, though carbon is more abundant. It's impossible to determine what's supposed to be smaller and what's supposed to be larger.
I never made these connections until now. This will really help me remember which is which, I honestly had just given up.
It's funny how different naming schemes "click" (or don't) with different people in different ways. When I first heard the haiku/sonnet/opus it was immediately obvious that they were linking the size of the model to the size of the piece of writing. Same with moon/planet/sun etc. for OpenAI's models.
If Google was going to go with elements, it should have been an absolute no-brainer to link size of of the element to the size of the model eg. Hydrogen/carbon/argon, but no, Google screws up the naming like they _always_ do. Hell, use the noble gases for christ sake. Helium/neon/argon/krypton/xeon would have been great, not perfect mind you, because a lot of people might not remember the ordering of those elements, but at least it would be consistent and make sense.
Yea, impossible to know whether it's supposed to go by atomic mass, frequency or some other metric. Plus most people off the top of their heads don't really have these stats memorized. With OpenAI and Anthropic it's way more straightforward. If you stop the average person on the street and told them to order the three by mass, most couldn't do it. But they could order Moon/Earth/Sun/Astra
Yeah. Atomic mass: Makes no sense, argon is larger than carbon. Frequency: Makes no sense, argon is least frequent of the three and hydrogen is everywhere, so carbon would be in-between rather than the flagship.
Letters of the alphabet: Also makes no sense. There was a theory that Argon = A, that there was an unreleased B model (Bismuth or Barium or something), Carbon = C, etc. (Similar to Intel's MMORPG-inspired Alchemist --> Battlemage --> Celestial...) Hydrogen broke this notion.
It's just completely incoherent.
Now written out like this, I can see a D = Deuterium, which is mass 2 or heavy hydrogen, but somewhere along along the line it was lost in translation.
I’m not sure if the average person on the street would know the terms Luna, Terra, Sol, and especially Astra.
Also, the biggest being Astra is a bit confusing. Sol is also a star.
Furthermore, unless you know they’re supposed to be ordered by size you can’t know which is best.
Corsa, Astra, Vectra and Omega would be better names.
I see that Kadett finally disintegrated into nothingness :)
The average person isn't supposed to care, they won't be switching models to prompt for the best television for under $1000. But for professionals they should have some sort of understanding of these things.
> I’m not sure if the average person on the street would know the terms Luna, Terra, Sol, and especially Astra.
I don't think these are as difficult as you're making them sound? I think the majority of people who will care what model they're using will know "lunar", "terrestrial", "solar", "astronomy" etc?
Google genuinely have made a huge mess with their AI product stack. They have Antigravity, Antigravity IDE, Gemini app, AI Studio, etc. They have too many similarly named and confusing products. Google should seriously fire their branding team and bring some new talent to consolidate and restructure things logically.
What's a fable? Why is it bigger than an Opus?
Maybe the AI is trying to tell us its real name !
Ar - C - H - ?
So the next model will be Yttrium?:)
If the next two are oxygen and nitrogen, which are plausible, that would be somewhat sinister.
when you put it this way, it sounds like they are operating on an XY axis (at least) vs a continuum. Argon big, complicated, slow AND expensive; Carbon abundant, widely available, can do anything, costly as a result; Hydrogen, abundant, volatile, cheap. can't imagine how insane Nitrogen must be
Sometimes I wonder if naming things is hard or what.
Of course Google is ordering their models by their density at standard temperature and pressure.
Hydrogen: this thing floats. Density less than air.
Argon: does not float. Density more than air.
Carbon: this thing is a solid, whichever allotrope you pick. It’s solidly heavier.
See, so easy! /s
I don't know if it's a good idea to market one's AIs by how "dense" they are.
Of course not. See the /s tag.
I find theirs to be a lot more excusable, given it's obvious that the Sun is bigger than the Earth, which is bigger than the Moon. Anthropic less-so, but if you know the meaning of each, you can infer it somewhat well.
Google's is just random.
Anthropic's naming does not make sense, but at least they are consistent.
What about it makes no sense? I find it to make the most sense, given it's about text generation and their models are named from shortest text to longest.
They aren't, though- a fable is only bigger than a sonnet. Technically, an opus is any creative work, including a music score. A magnum opus would be a person's most celebrated work, but that doesn't have to mean biggest or longest.
I guess it's poetic but I don't see how haiku < sonnet < opus < fable < mythos is unclear. If you've never been exposed to literature maybe?
You day that very matter of factly as if it doesn't include "opus < fable"
I think people on HN argue way too much about this. It's not like they're named iPhone 18 Pro XL Max. It's meant to evoke a feeling of wonder or awe, not really that hard of a logical leap to make, and it's only one minor quibble. Jeesh.
Mythos (Greek) = Fable (Latin)
Fable is essentially Mythos but with guard rails added.
I'm still not sure if Carbon is meant to replace Argon or meant to be used alongside it?
For Google Workspace users: please replace the ancient 3.1 Pro in Gemini. I think it's 10 months old...
The general trend -- naming models classes after objects from some implicitly ordered set -- was started by Anthropic. It's generally fine, but they didn't leave enough headroom for their largest models, so the implied ordering for opus/fable/mythos is kind of muddled to the average person.
Small/medium/large/etc. has the most easily deducible ordering, but it's not great from a marketing perspective and also has headroom issues unless you're ok with names like gpt-8-xxxxl.
OpenAI's naming is probably the best. The ordering is easily deducible and there's plenty of headroom in both directions. Tons of cool sci-fi sounding names still on the table, and it also comes with the implication that they're approaching a singularity.
I don't know what Google is doing. It's kind of impressive to have an inconsistent ordering after only three names.
OpenAI's mostly works, except Luna and Terra are swapped depending on whether you think of it as "distance from earth" or "size/magnitude of object".
Google, though, really has broken new ground in incoherence. WTF.
Naming an AI model Argon will never not be hilarious to me.
It doesn't do anything! It's inert!
I'm waiting for Gemini 5 Uranium.
Which will be launched after GPT 7 Uranus.
Same, my eyebrows about jumped off my face. To go from Google as normal -- product announced, rolls out randomly in some access modalities, mysteriously doesn't in others because no org talks to any other org -- to "wait wtf is this new model I didn't even hear about this" is so laughably bizarre. Not going to get used to it, I know Google.
the naming scheme "trend" (opus/sonnet/haiku vs luna/terra/sol/astra) is just fine. It's just Google's version of it that sucks.
Agreed on the naming issue and while it was novel at first it will inevitably get worse as more models get release, until decide to abandon it all together. It’s like they are setting themselves up for failure on that front lol
Look at how many AI adjacent products Google talks about in this press release along with its model names. Anthropic and OpenAI are guilty of it as well, so many models and ways of working with them right now, letting the users figure out the nuances of each.
Earlier today I found myself asking ChatGPT to research and explain exactly what Anthropic Playground is for and how specifically it may differ for my tasks from Anthropic Workspace, Chat, Code, and API credits. I chose ChatGPT Sol 6.1 high for this task going on gut feel, hoping it was the best model to handle that job without wasting my precious Claude tokens.
I think there need to build these adjacent things to create lock in. I can switch between Claude, GPT and pretty much and model on OpenRouter in little time, especially for dev work. With all of these being weeks apart in progress on the models and open-weight models seemingly getting closer and closer there is almost no moat. Without a moat they cannot charge high prices - it's a commodity. Having lots of my data in their system and having other exclusinve features like Claude Design or Whiteboard and what not, they create something that I cannot keep if I switch to a different vendor for a cheaper model.
Yes the naming is nonsensical, because the intelligence ranking is this (with atomic numbers in parentheses for those that don't have the periodic table memorized):
Carbon (6) > Hydrogen (1) > Argon (18).
Double nonsense because it appears Hydrogen is in fact a smaller model than Argon, in which case this should obviously be released as "Gemini 4.1 Hydrogen" or something.
What they probably should've done if they really wanted to use elements is to do noble gases, and start these first three models in the middle (to leave room for bigger/smaller models in the future).
So you'd have:
Argon => Argon (18)
Hydrogen => Krypton (36)
Carbon => Xenon (54)
Not exactly because as noted it appears hydrogen is in fact the smaller model, but then this should be (again) 4.1 or something.
And they'd leave space for a truly massive model if needed (Radon) and two slots for much smaller models, with Helium (2) presumably being a model that could be run on basically any consumer device.
Maybe the confusion is the plan; keep people talking.
No matter what you do you can't go wrong if you spin it right!
cargo culting
"Claude and Codex have cool names, they are more successful than us, so therefore if we give gemini cool names we'll be successful too"
-marketing drone, probably
they probably went to chatGPT and asked for good names
As this is a) being served in Search, b) supposed to be 2.5x faster than 3.8 Flash, c) the element with lowest atomic weight; I deduce that, despite being more capable than 3.8 Flash, now the “entry level” model set to replace 3.5 Flash Lite.
But it is priced the same as the non-discounted 3.8 Flash, considerably higher than what anyone is paying today for 3.8 Flash or 3.5 Flash Lite or the deprecated 3.0/2.x Gemini models. How does this work? Is Google abandoning the “fast and cheap” tier? That seems like a mistake since plenty of workflows are absolutely price inelastic, meaning when the 3.8 Flash discount ends at EOY, lots of current enterprise Gemini customers would have to take their business to another shop??
Same impression here. If this is Google's strategy we'll have to migrate to Luna, because costs are already high for our workload on 3.8 Flash (which in itself was already a 40% increase compared to 3.7 flash due to being more verbose and thinking more) and this now doubles cost once again.
Yeah, and they rug-pulled 3.7 very quickly, silently redirecting to 3.8 Flash. Raised our prices and broke some workflows.
Yeah they fuckin did
Isn't something open weight cheaper (and probably better as well)?
Depends on your needs. Gemini Flash has best bang for your buck on visual tasks and best multilingual support/intelligence for non-Western/non-East-Asian languages; better than paid or OSS alternatives.
Yes I gave for granted textual tasks but they're not the only ones. If I remember correctly, mistral large should be good as well on image tasks while being cheap(er). I never had to do anything with less common languages tho
According to Artificial Analysis, it's much cheaper per task than 3.8 flash, 3$ for H xhigh vs 6-7$ for 3.8 medium-high. I don't understand how though, since AA says it has more output tokens per task.
Hi there - AA employee here. This is due to 3.8 Flash having a much lower cache hit rate compared to Gemini 4 Hydrogen
Why would cache rates differ? Does hydrogen have a longer cache period?
sorry if this is a dumb question, but how is this possible? what can cause the cache hit rate to change just by changing the model?
This is based on what we measure from the API - we see different endpoints depending on how they are served can have different cache hit rates (cache hit rate is defined as the % of cacheable tokens that actually come back as cache hits)
In the Gemini app it's replaced Flash. Lite is still there.
But that's an old model, and if they replace it they have no smaller element to use for it's name. I guess it would have to be Hydrogen Light.
Quark/Neutrino/Axion/Tachyon.
I suppose they could name it proton or hydron.
The fact that there is still a flash lite dropdown in Gemini points to them keeping the flash lite model family.
Isn't Flash-lite or related model the one used for AI summary in Google Search?
So all models will be named after elements except Flash Light? Or maybe the next version will be named Hydrogen Light?
Gemini Photon?
Gemini Electron?
> considerably higher than what anyone is paying today for 3.8 Flash or 3.5 Flash Lite or the deprecated 3.0/2.x Gemini models
It's cheaper than the last Pro models.
Anyone else got suspended for just using agy on the command line? "This service has been disabled in this account for violation of Terms of Service. Please submit an appeal to continue using this product.". The support claims that it can do nothing, not even see a reason and pointed me toward a form with no text inputs just a checkbox along the lines of "I understand the TOS, please check". So I filed an appeal, and it seems that it was automatically processed and within a few minutes. agy was working for one more prompt and it triggered a suspension again. What channel can I use to get any attention on something like this? I've heard these Google horror stories... fortunately this is only the Antigravity part, but I've just today upgraded to Ultra and this coincided with that. I've become extremely hesitant to continue using broader Google services when this can not be resolved via their customer support.
> Anyone else got suspended for just using agy on the command line?
So I'm a little spooked by this; if you don't mind, can you be specific about exactly what you did that seems to have preceded this?
I am also spooked. I reached out to support once more, and they seem to have escalated this to backend support (seemingly working on a fix live?). So this seems to be some larger issue and not a "typical" false positive. If I get to know anything about the reason, I'll follow up here.
The thing that preceded this was an upgrade to ultra today from pro, and a call of agy cli from within opencode (explicitly not using it via some login/oauth, but just calling `agp -p "prompt"`, which had been working for months). Right now it seems that every agy chat works for 1-3 messages, then gives a suspension error.
If this gets resolved via live/email support timely I would be fine. But I am spooked by the appeal form/process that is completely opaque via a google form(!) and seemingly can can take days: https://docs.google.com/forms/d/e/1FAIpQLScOJWibQ_-hYuVv63kJ...
I used agy for apps and had no problems... However, the idea of this happening never occurred to me. It might make more sense to have my AI services under a different account :-(
The slop is strong with this one
If there are any Google employees reading:
I want to go all in the Gemini ecosystem but several things keep me going back to OpenAI:
* the Gemini app is glitchy on iPhone (my chats frequently fail or are lost if I background the app)
* the Work equivalent (I guess task?) is not as powerful. I often have ChatGPT Work to arbitrary computation and even install ML on the short-lived VM.
* there is no simple integration between antigravity and the Gemini app. I want a way to continue working away from my server while using the familiar app. No I don’t want VNC.
* ChatGPT sites are cool for putting up proof of concepts with zero effort
* Getting banned from Gemini could mean nuking your gmail account. Can't afford that risk even at low levels.
I believe this has been rectified and now if you get caught breaking Gemini ToS, it just bans you from Gemini (I think you might even get a cool-down warning ban first).
For more egregious violations like the kid uploading naked pictures of himself, I don't know.
They said it was rectified and some weeks later people kept getting banned.
You can't trust them. It's not worth the risk. And the models and lack of subsidy definitely don't warrant the risk.
Glad to see this. However, given the general level of product confusion in the Gemini lineup, I'm not planning on relying on this even if it is their intent.
> there is no simple integration between antigravity and the Gemini app. I want a way to continue working away from my server while using the familiar app.
On the other hand, I'm very glad they have separate quotas.
FYI, you can use Antigravity as a web app with "remote control" mode. It works pretty well, better than VNC since the page is responsive, though the text is pretty small and the UX if not really differen than the computer version.
I've found agy remote control doesn't work anywhere as nicely as claude code. I tried using agy on my PC, enabled remote control, then opened on my Android browser. If I type anything, there is no way for me to delete (backspace) if I make any typos other than refreshing the page. It basically made the experience unusable for me. I don't understand how something so basic as the backspace button can be broken. Every time I try a Google AI product something very basic is broken. PDF uploads on the Gemini app also don't work nearly as well as with ChatGPT or Claude. These are all things I'm sure Gemini Carbon can find and fix in 1 day. I don't understand how they ignore basic UX to such an extreme.
If I had to guess the backspace bug is usually the Android keyboard. Gboard sends composition events instead of keydown events, and a lot of web UIs only listen for keydown. Switching off predictive text or trying a different keyboard fixed it for me.
Otherwise you can use agy on the Cosyra mobile app, its far better. I'm not going to claim its better than the Claude app, but its certainly better than the Gemini RC.
Does the antigravity iOS app give you the remote coding features you're looking for?
Is there an antigravity iOS app? I just searched on the App Store and couldn’t find one
Sorry, no there isn’t - I thought there was but it’s just a pwa for https://antigravity.google.com/
I've only briefly tried it, but can (at least on Linux), run `agy remote-control start` and then access the device through antigravity.goole.com
I think a new default free model for AI Mode being so close to the frontier will be a massive jump in perceived AI intelligence by the general public.
It will be the same as these smaller models from OAI and Ant. Strong in benchmarks but pretty bad in real world use.
I've read this same comment every time a new Luna came out, and the difference between free GPT (Luna) and paid (Sol and Astra) is still outrageous for general use.
I think this is a completely different class from Luna and is much more akin to 6.1 Sol in terms of capabilities and general intelligence.
Maybe, although google tends to benchmaxx. If it's actually sol for free that's kinda crazy good.
3.8 flash was easily the best general everyday person model yet though
3.8 flash was never served to free users on google though?
3.8 Flash was served to free users in AI Mode IIRC
Only for pro and ultra users
The free default was 3.5 flash lite.
Ah, you're correct.
Maybe I’m spoiled but I found it repulsively dumb. Some of the highest rates of hallucinations since GPT-3. Incredibly lazy. Heavily biased to parametric knowledge and speed over accuracy. I would rather have no AI at all than use that.
It's definitely not the case that small models only do well in benchmarks. We're working on a battle system on top of OpenCode, so have data for real people doign real work.
Right now DeepSeek Flash 4.1 Flash has a 64% win rate against GPT 6.1 Sol. Small data still, but it has 33 wins vs. 18 losses, which is confidence interval of 52% - 78% win rate. https://agentduel.com/leaderboard/models/deepseek/deepseek-v...
That's coding though. I was talking about general usage.
Okay, makes sense, though I don’t see a huge difference in quality between Luna and Astra for basic chat usage.
This is the most important class of models to keep watch. These are the ones that are going be to workhorses in the post-tokenmaxxing clarity
Disagree, Opus-tier models are the workhorses. You can get a ton done with them on a $100 subscription.
And for those on API billing? Opus is still quite pricy for what you're getting done with it.
> And for those on API billing?
Opus-tier on open models and local hardware.
If you're on API billing then presumably you're a business, and $1,000 a month is still reasonable for that quality of work
Those subscriptions won't be subsidized forever. At least not as much as they are today.
API token inference is subsidizing the rest of the business.
People keep saying this, but inference itself is not what is being subsidized. The expense for electricity/hardware/datacenter operations/etc for an inference token is less than what they charge for it, so that cost doesn’t HAVE to go up to be profitable.
The thing that is being subsidized is the TRAINING costs, and that is because they aren’t selling enough tokens yet to pay for the training costs. This is just like most other growth businesses, though, and if enough people use enough inference, it can be self sustaining.
Price per token is dropping (though anthropic are doing dastardly things such as using new tokenizers that use 30% more tokens), efficiency will likely increase.
If the subscriptions (of all of the big companies - unlikely?) give a poor amount of value in the future, will move to trusted/open router and use a harness that uses the latest cheap good models for most things, more expensive models as orchestrators.
Already there are non-SOTA models that should be sufficient for most non-coding tasks, and soon for coding, also.
The future is bright, in that respect, without subsidizing the ridiculous IPOs.
Or $20 for the same amount of work with US/EU hosted Deepseek flash v4.1.
Why stay on API billing? A $200 monthly subscription has a crazy amount of usage and more than enough for a couple of concurrent projects.
Because you can't get subscription if your company has more than 150 users.
So they move to non SOTA models.
Enterprises with more than 150 seats have to pay API costs.
Or get multiple teams. Just use git instead of artifacts ;-)
For users who pay subscription costs, 100% right.
But for enterprises, who have to pay API costs with no discounts, no way.
Assuming the price would converge to the subscription price, not the API price.
I thought Opus is strictly worse than Fable now?
Only because you're being heavily subsidized.
They make money on API costs not subscriptions. Subscriptions are not going to last forever
I like that you gave an argument and careful reasoning instead of just a vague claim based on common meme about ejaculation.
How is tokenmaxxing like ejaculating exactly? I don't follow the analogy.
Here is the argument
1) There was a false narrative called tokenmaxxing whereby one is more productive by spending more tokens
2) This narrative was fueled by discount prices on tokens (which still happens to this day on subscriptions)
3) There is a point where we collectively realize that tokenmaxxing was a state of mind that is not rational for either the consumer or provider
4) We will then pay API prices because LLMs are obviously amazing but no one will offer subscriptions
5) API costs of Opus 5.5 and Fable tier become uneconomical
6) That leaves us in Sonnet/Haiku tier
I understand his assertion, and I can make up an argument that fits it too, probably a few arguments that aren't even compatible with each other, but what you said isn't what he said, you are just assuming all of this, and also nowhere here does it explain how this fits an analogy of "post nut clarity".
Why is this exchange reminding me of that You Can See Everything documentary?
The analogy made sense to me: Tokenmaxxing policies were the result of an irrational drive by execs who just craved an, ahem, explosion of metrics due to the excitement over new technology. Now that cost sensitivities are setting in, there’s a growing sense of unease that maybe they were in an altered state when they created the mess and now that the excitement has passed they have to clean it up.
So people had this drive to use tokens, and once they experienced the release from using the tokens they went into a refractory period where their reasoning is no longer being overridden by the need to use tokens? And later I guess they'll start to feel more and more of a need to use tokens and this will repeat?
Or is it just a throwaway phrase that makes no sense?
We’ll see if the cycle holds true for tokens or LLMs specifically, but for sure there is a larger “new sexy tech” cycle with a strongly emotional output===outcomes metrics irrationality phase to it.
I think he's just saying tokenmaxxing is a temporary hype lol, probably fueled by subsidized subscriptions. When that's over, these cheaper models might get more usage.
Can anyone understand the logic behind the names? Argon, hydrogen, carbon, what is the mental model one should use to understand where they fit, or what comes before/after?
Haiku, sonnet, opus.. is ok
terra, Luna, Sol... is also ok..
but this?
They started out alphabetical for their pro releases: argon, barium (never released), carbon
Then they needed a flash model so they picked the lightest element.
They dropped the ball pretty quickly
Google has a chronic forgetting-about-their-products problem.
The higher the atomic mass, the higher the intelligence.
It's not, Gemini 4 Carbon is better than Gemini 4 Argon.
Argon internally was called Barium (Argon was a different model) so the models were originally named in alphabetical order. Then they change their mind and decided to call the lightweight model following the atomic weight so Hydrogen after they have released (A)rgon and (C)arbon. I imagine that this lightweight model is the original Argon.
H lightest (highest speed), Ar stabilizer (balanced), C structural bonds (highest reasoning)?
I like the theory, curious how this will evolve though
Carbon > Argon > Hydrogen not that hard?
It's 3 items, it's not difficult. But there's no way to infer that from the names alone and that's Bad Product Naming.
Also... Argon and Carbon are the same price, but Carbon is better I guess?
And "Hydrogen surpasses Argon on the Artificial Analysis Intelligence Index at its default high thinking level"
Given it's twice the price of 3.8 Flash I hope they would find room for another cheaper tier, but the Hydrogen name suggests not.
I'm glad Google are coming out with some decent models but it's a slightly confusing line up.
Always rely on Google to mess up product launches
They are all terribly named, but Gemini models definitely take the cake right now.
My guess, by atomic weight. Hydrogen = "sonnet". Carbon = "opus". Argon = "Fable" only argon 4 just happens to be kinda shit? But also it's google so who the hell knows. They could be just going alphabetical argon, then carbon, then hydrogen as release order follows the alphabet. That'd be kinda dumb but so is their access/subscription strategy so it'd be on brand.
Reasoning-wise Carbon > Argon (so actually is Carbon, Argon = "Fable", "Opus").
Carbon can be more performant while being a smaller model.
I thought they were going by alphabetical order like intel does for GPUs
A = argon for initial release B was skipped because you can't name a product boron C = carbon D = ??????
They might be going for atomic weight, but the average person couldn't say whether argon or carbon was heavier.
"Argon" had me thinking they would go with a noble gas theme, with Helium as the light one and Xeon as the heavyweight.
Personally I found my usage of llms bimodal, I only have uses for the top frontier models (for research and development) or for the cheapest and fastest ones (to use in production for consolidated, simple flows). I am curious what people find the mid tier models useful for.
I feel like people use medium models thinking they are saving money, but in the end, it ends up wasting more time and tokens theough the increased rate of issues/bugs.
If the highest tier model makes even one or two fewer mistakes that require human intervantion, that almost makes it a no-brainer choice for production.
I've been doing some research lately on my own (aside from the major benchmarking going on) and I've definitely found cases where medium models do much poorly and cost more than the higher models. But on the other hand, if I can give the medium models certain skills or other resources to complete a task then they perform better in cost than the higher model (which, given the same resources, sometimes gets more confused than the medium does).
At least for my use case of evaluations of a set of features.
Yup, in my benchmarks usually medium or high do better than xhigh or max
You’re talking about something else (reasoning levels). The parent comment is referring to e.g. GPT Terra as opposed to Luna, Sol or Astra, Claude Sonnet as opposed to Haiku, Opus or Fable, or Gemini Flash/Hydrogen as opposed to Flash-Lite, Argon or Carbon.
Oh, good catch. I do use Sol in production, for cases where I need quite good results, but can't afford to run in production the highest tier model due to pricing (so it's like the highest tier model I can afford for that use-case).
To be more specific, this is for a support automation, where the model has to be quite accurate in following rules and generate responses, and Luna models or lower-tier models often make bad mistakes and generate wrong responses, so are not really usable for sending data directly to the user.
Even better at the local LLM stage, where QFN can catch a mistake in prior work and patch it without any subscription costs at all. then I can use a cloud model for polish and finish and move on.
I'm building stuff at home with my work focused on managing infra. If anyone wants to sponsor a SOTA model subscription I will check it out. Until then I'll see how flexible a Strix Halo setup and my homelab can be.
Asking people to guess beforehand how much intelligence they require for a task will hopefully be a relic soon. We’re lacking useful routing techniques which accurately pick the right model for the job. Or a high intelligence model which completes simple tasks fast and cheap. Or something else.
Yeah, an xhigh models should know a task is easy and respond quicker.
I have some tasks that deserve a bit more reasoning but for a long time most of my needs have been met by mid-range models, rather than the big ones. There's barely any motivation for me these days to use something more advanced than Terra on low/medium for everyday work. For a personal agent doing occasional coding, I'm still on Mimo flash.
I think many people still expect their tasks to need something more fancy, but could do with a reset of expectations and try to go much lower.
Same, I either use xhigh on best model, or use cheapest and fastest models (mercury/jev/haiku/luna)
In the Opus 4.6 days xhigh was a meaningful improvement. These days with Opus 5.5 I just don’t see a meaningful quality increase.
In executing tasks no, but in generating assets, art, UI, etc, higher reasoning leads to drastic improvements.
Production pipelines for non-coding usage. Jagged edge is very real. Just one example, loads of tasks where Gemma 4 26ba4b wipes the floor with Deepseek Flash v4.1. On Artificial Analysis-like benchmarks, the latter is many times more "intelligent". Not everything in the world is coding and math, but those are easy to "measure". Things that are hard to measure aren't part of their benchmarks.
The pricing on these models make really little sense. This is $1.5. Argon is $2. Carbon is also $2 for now.
Feel like this pricing only makes sense for a product where you don't even select the model. Maybe Antigravity is like that.
Isn't $2 the intro price for Argon and Carbon?
Do you know if Argon and Carbon are also priced the same inside subscriptions?
It's starting to feel like OpenAI is in 3rd place now. Really eager to see how / if they can claw their way back up.
All the labs are holding back releases now, not necessarily to play games, but to make sure alignment is right. These models are teetering on the edge of "One person can seriously damage digital infrastructure" and an event like "Teenage kid uses Claude to bring Wells Fargo to it's knees for 6 hours" would be absolutely catastrophic for the whole industry.
> Teenage kid uses Claude to bring Wells Fargo to it's knees for 6 hours" would be absolutely catastrophic for the whole industry.
It would (rightfully) be catastrophic for wells fargo. And great for the industry. If your security model relies on children not having access to things state actors can leverage, your in for a rude awakening.
Right now it's looking like the security model for everyone is "Does or doesn't have access to SOTA LLMs"
Hence why they are gating model releases. Even Chinese models are doing this as Kimi K3 shipped with a lobotomized cyber sec ability. The major labs all have invite only cyber models too.
> Does or doesn't have access to SOTA LLMs
“Security by availability” doesn’t seem really different to me from “security by obscurity” which has already been proven to be complete bullshit.
Then you are in for one too because your personal security relies on that, guaranteed. So does everyone's.
And the sooner that goes away, the happer I am. My confiendence in the orange dude, Winnie the pooh and their friends is honestly close to 0.
Wait 1 week top, they'll release something soon for sure
Oh I have no doubt they're working hard. It's just google is going to strangle them on the integration layer if they're not careful.
And hopefully this time for them they don't have to release a new rev the week after.
I wonder now that Terra-6 is out how it would compare to Sol-6. (Which was replaced a week latter by sol-6.1 after chat got dropped opus 5.5 to incredible fanfare then sonnet 5.5 that radically upped the medium model ability).
I would expect a new release from them soon to climb back up.
Heavy user of all models here and Astra is million miles ahead of everything.
I have a stupid benchmark: Bongard Problem 2 from here (https://matheminutes.blogspot.com/2019/04/bongard-eggs.html). It's an open-ended visual task that is easy enough for humans, but models are still somewhat bad at visuals, so it's one of the few remaining reliable ways to trigger hallucinations. This one answered correctly, but then felt the need to add a fun fact that was wrong:
> The rule separating the two baskets is layering (which color is on top). Left basket: In every pattern on all six eggs, the blue shape is on top of (overlapping) the green shape. Right basket: In every pattern on all six eggs, the green shape is on top of (overlapping) the blue shape. [true solution]
> (Note: Both baskets contain the exact same distribution of motif counts per egg—two eggs with 2 pairs, three eggs with 3 pairs, and one egg with 4 pairs—serving as a classic Bongard distractor to hide the occlusion rule.) [factually incorrect and also would make no sense as a distractor]
And then trying a harder puzzle just makes it bullshit completely.
I don't understand why Gemini models always add these "it's typical XYZ" bits, which are typically so much worse than everything else in the answer.
The sample problem I kind of gave up on and guessed sort of... it couldn't be sorted by color that's too obvious.
And then the real problem number 1. The answer is "Left - each figure contains an even number of arrow heads. Right - each figure contains an odd number of arrow heads"... but on the right I can clearly see that the three eggs across the top have an even number of arrowheads.
It's not the number of arrowheads on the egg; it's the number of arrowheads on each of the little figures on the egg. Those eggs you mention have 2 figures each, true, but each figure has just 1 arrowhead each.
If you have an Ultra subscription (I don't or I would try it) I would be really curious if Carbon is able to solve your harder Bongard problem without making stuff up given it's absurdly high AA Omniscience non-hallucination score, it's supposed to be like 30%+ better than nearest non-Google competitor. Although Hydrogen xhigh apparently scored within 5% of Carbon so who knows ¯\_(ツ)_/¯
I don't have that subscription, but for the record, it's this problem: https://imgur.com/a/idyH0Kh
I won't share the solution (you have to keep some mystery in the world), but I showed this a year ago at a high-school reunion and five people independently found the same solution, so it should be well-posed.
Ha, okay I caved and spent the 5¢ to ask Carbon the problem through openrouter and it was indeed able to solve it. No random hallucinated fact either. Thought you might want to know. Cool problem!
Wow. On paper this looks like the best Gemini release OAT. They really dialed in the usage and cost issues.
On the AA benchmark it kinda makes Sonnet and Terra look weak. Hopefully that translates to real world usage. Can’t wait to try it out.
I was really hoping they would change their mind about the increased pricing for Gemini 3.8 starting new year, but apparently it's set now
Good they are able to replace Flash. And cool this is available to everyone via search. But I still don’t understand Googles strategy with these new models.
For pro sub, Argon is still very limited and Carbon is only available on Ultra. This is not competitive at all. To say nothing of also being locked into Antigravity.
I'm very impressed with it as a model for image->html builds. It's the first non-anthropic low-param model that successfully captures complex flows.
Comparisons:
Hydrogen: https://html.non.io/lcars-gemini-4-hydrogen
Sonnet 5.5: https://html.non.io/lcars-sonnet-5.5/
Original designs: https://diffui.ai/app/canvas/5093e689-1e74-4f26-b632-2a4500f...
IMO it sits between sonnet and opus in terms of quality, and is equivalent with gpt sol (for this particular benchmark). Google did a great job with this.
How do you actually use hydrogen in the app? The only model selection options I see are "auto," "code and analysis Argon 4" which is paid, and 3 thinking levels.
Nice upgrade from 3.8 flash.
My use case benchmarks 37% more expensive per task but 40-60% of the total time per task (median and p90 had the same improvement).
Both models were smart enough to complete all eval tasks, speed was the main thing we were struggling with. Both models on medium reasoning.
Overall feels like a good upgrade.
The orchestrator in Antigravity might be the most important part of this release. I'm excited that this is being built natively into harnesses.
I really hope hydrogen keeps the excellent conversational ability of 3.8 flash. For brainstorming or even decoding the dense output of the powerhouse models, flash was a real delight to use. Its the first non-SOTA model that I would actually go to often over the larger models, it was fast, plenty smart enough, and generally wrote really well. Here's to hoping hydrogen keeps that up
In the Gemini app of over the API? In the app Ive found it to write poorly, probably a system prompt thing. I've done some stuff with Argon over the API and it wrote well.
This is a good blog post because of the graphs comparing directly to Terra and Sonnet for both wall-clock time and cost per task. This is the exact information I'm looking for.
My TL;DR is: seems like a better choice than Sonnet in areas outside of coding, and at high reasoning it's smarter than Terra but somewhat more expensive.
Why all the Ai companies are dropping the names that meant something like flash, mini, pro to these meaningless name. It's very confusing
Google's AI product stack was always confusing and this just adds to it.
Can anyone compare usage allowance on Google subscriptions vs Claude or Codex?
I'm really enjoying all these Gemini releases. I know a lot of people keep saying Google is too far behind in the race and that they dropped the ball to early but I don't feel that way at all. I had both Opus and Fable try to find a discreet bug in my codebase and both models assured me I had nothing to worry about and was probably overthinking. Argon was the only model that got to the bottom of the matter and developed the fix for me. I constantly find the Gemini models perform better in the real world as compared to the other frontier models and I also find they have less of the AI-isms that many of us have now come to associate with ChatGPT and Claude. Gemini feels like you're speaking to a comrade. Claude feels like you're speaking to a friend and it's an important distinction I feel.
Qwen 4 Pro preview has been beating fable-5.2 in my tasks. This model is available to try on the free tier. I can't even use argon or carbon on my $10 google sub. Why doesn't google show off their top models with a few free prompts a day to get us hooked? Let the dog see the rabbit.
Have you tried Antigravity? You should be able to get a chance to play with it for free in there I think.
Yeah, I used to like it, until they dropped pro models. I am on the AI plus tier now and Argon and Carbon are not included.
Oh hey my google workspace account actually got this. It was on like 3.6 and 3.1 for a long time.
Google are doing such a terrible job. They aren't allowing Pro accounts access to this in antigravity that they gave to students or via mobile plans elsewhere, because they messed up badly and don't have the compute. They have been absolutely horrible about communicating this and even gaslighting the users, completely nerfing the same models that they do get to use, without telling them that they are doing.
So after being caught up in this, it's left a very bad taste and I'm subscribing elsewhere.
I went to geminiIsland[1] and asked for spongebob's, squidward's, and patrick's houses [1] https://geminisland.raph.plus/
These names are honestly too confusing. Given carbon, hydrogen, and argon I have no clue from the names which is small, medium, and large
Gemini was never my first choice. The only thing built on Google models I use is NotebookLM.
Where is Gemma5?
Where is Gemma5?
Where is Gemma5?
Where is Gemma5?
The naming of these Gemini 4 models is ridiculous.
In order of sophistication and cost, the models are Hydrogen -> Argon -> Carbon.
What kind of ordering is that?!?
With the introduction of Hydrogen, it's no longer ordered alphabetically. That would be Argon → Carbon → Hydrogen.
And it never was ordered by atomic mass, which is arguably the best ordering if you're going to name your model classes after chemical elements. That would be Hydrogen → Carbon → Argon.
It's also not ordered by prevalence in the universe. That ordering would also be Hydrogen → Carbon → Argon.
It's not ordered by electronegativity, boiling point, melting point, or any other remotely defensible chemical property I can think of.
It's not even ordered by letter count, Scrabble score, or NFPA 704 Flammability Rating, all of which would be Argon → Carbon → Hydrogen.
Per Business Insider[0], Argon and Carbon were part of an alphabetical lineup:
> An internal document shows that Google has tested a series of Gemini 4 models, internally named Argon, Barium, and Carbon.
Then I suppose they tacked on Hydrogen since it's supposed to be a "light" model and hydrogen is the lightest element. I agree that it's super confusing though.
[0] https://www.businessinsider.com/google-employees-test-new-ge...
Density at standard temperature and pressure? Carbon admittedly has a number of forms at STP but they're all solid and thus heavier than argon gas.
Density of the most common (on Earth) allotropes at STP, perhaps?
Or maybe they just pick them out of a hat.
nodejs release names
alphabetical skipping the weird names
next would be iron krypton and lithium
+1 to this the naming is totally opaque and ridiculous.
There is no hint of what each one does. I get that these were internal code names, fine, but to everyone outside Google it's not helpful to have these meaningless names.
ChatGPT and Claude aren't much better - there is still quite a lot of assumed knowledge there assuming you know what e.g. a sonnet is and what an opus is - but at least if you squint there is something there.
Google's names just seem to be randomly generated.
Maybe they were worried that the 3.1 Pro model being around for so lunch ruined the brand? But 3.8 Flash was pretty decent I thought so it's not like "flash" was burnt. I dunno.
The choice is really odd. Hydrogen and Carbon could have an organic chemistry theme going if they chose Oxygen, but then they go with Argon? Anthropic and Open AI have the names figured out in terms of obvious size: Luna->Terra->Sol->Astra and Haiku->Sonnet->Opus->Fable->Mythos makes intuitive sense. I have no idea how the Google model size class works.
I know some people are still a bit confused by astra, sol, terra, and luna, but at least you can explain the reasoning and it makes sense. Google's new naming scheme is just pure luna...cy.
I don't know why but this post reminded me of the show(books) Altered Carbon.
Also totally agree. What a bad way to name models...
We'll know they've achieved super intelligence when tech companies suddenly have sensible product names.
Yeah, they really fumbled the naming scheme for this batch of models.
At this point even using Starbucks sizes would be better.
I now believe that they're trolling.
Surprisingly plain pelicans for a Gemini model, quite a nice illustration of how Hydrogen really is significantly less powerful than Carbon and Argon: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
For comparison here's Argon: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
And Carbon: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
Here's a comparison grid: https://static.simonwillison.net/static/2026/gemini-4-pelica... - note that Carbon and Argon are priced the same, so it's actually quite neat to see the difference in token usage between them (Google say that Argon is more token efficient.)
Their naming scheme seems a bit tough to follow. But it does seem like they finally learned their lesson and announced something at the same time it's actually available!
When are getting more powerful Gemini 4 Lead, and Gemini 4 Uranium models?
Gemini 4 Plutonium will have nerfing half life?
Looks like they are releasing these... periodically!
3.8 Flash pricing will double in January to match 4 Hydrogen pricing
https://aistudio.google.com/docs/pricing#gemini-3.8-flash
First interaction, me: "Hello how are you"
Gemini: "I am operating normally."
Smooth.
I prefer computers that talk like completed, but I get I'm the minority in this.
Argon, Carbon, now Hydrogen. If Google keeps climbing the periodic table, the roadmap writes itself: Californium finds gold in your codebase (enterprise only)
So a pretty mediocre model at best on par with the open Chinese models. At least the Chinese models have a coherent naming
By what metric?
The naming of this stuff is so bad. Hydrogen (1) is dumber than Carbon (6), sure. But Carbon (6) is smarter than Argon (18)!
The labs use a pretty clear progression: Luna -> Terra -> Sol -> Astra; Haiku -> Sonnet -> Opus -> Fable/Mythos. Google doesn't make it easy.
Claude is the same really. Fable is a short story, so makes no sense in this hierarchy, and Opus have to think it as derived from "Magnum Opus" to make sense, else Opus by itself simply is a piece of music by a composer and would've been more appropriate for low-reasoning model to be used by sub-agents.
Anthropic, OpenAI and literally everyone else's naming schemes are in the same class of silly. The regurgitated complaints on every single thread are really getting tired.
I get that hydrogen is the smallest element but like it's not exactly that intuitive for everyday people to know which to choose. Whatever happened to big/medium/small
The era of "ai le bad" jokes might just be over
Billions of people will now be exposed to a model that crushes sonnet 5.5 on benchmarks (in google ai search suggestions).
I kinda miss dumb ai in a way actually, was a certain charm to it
Next will be gemini 4 cornholio if they keep following this naming
Hydrogen... the smallest atom that can still call itself an atom... and they've named their model after it?
I do not understand ai marketing.
looking as the price, seems like google wants to share the AI bowl with other companies, so generous!
though I suspect they keep the price this high to increase the profit margin for the shortened release cycle (instead of 6 months I think they will release newer checkpoints in 3 months or less from now on).
still, my application now permanently use deepseek models instead. no point using gemini, neglectable real life improvements for extremely higher price.
Pretty big Gemini fan, but my initial results with Hydrogen and Carbon are disappointing.
They are trading improvements on pass@1 pointless benchmarks for worse disasters at p90/99 that you're going to hit every week in agentic workflows (if not daily):
overreach (doing more than asked), overwork (when it should have failed fast), harm (doing the opposite of what was asked), orthogonal work (doing something completely unrelated to what was asked), etc.
Google doesn't seem to be the only lab making this trade...
The set of benchmarks they released are disappointing. 10 mostly random benchmarks. Really?
No one on HN seems to get that agentic workflows are not the only name in the game. Everyone doing agentic development is either on a plan or with a custom enterprise offering, either way using higher-performing models. These entry level models are (also) for building apps on top of while you are paying API prices.
There was this almost peaceful time few days where antigraivty drop down menu was
argon
flash
opus
sonnet
now its
argon
hydrogen
flash
opus
sonnet
gptoss (i swear ths was removed earlier)
carbon
oh ahd hydrogen has 4th extra thinking.
holy shit just route this shit for me..
E: or not
This service has been disabled in this account for violation of Terms of Service. Please submit an appeal to continue using this product.
Haven't done anything, worked fine before update, now after. And lol where is even appeal form. Right after I said Google One was good value.
Terrible naming convention on the low-end, because what do you do if you make something smaller than Hydrogen next? Gemini 5 Electron?
On the high-end, I feel like there is an implicit promise that if they keep releasing larger models and go past Oganesson (118), they are on the hook for AI-discovering a new stable element, so that is fair at least.
Just do the same as other products...after a couple years you wrap round back to the original naming because by then people have forgotten it
Honestly I cannot keep up with the naming conventions that keep changing every month.
Argon, Carbon, Hydrogen?
Luna, Terra, Sol, Astra, Betelgeuse?
>For everyone: Start using it in the Gemini app model drop down
and then it says for free users there is only "automatic" and you will get Gemini 4 Hydrogen "when it's the appropriate model for the prompt". Yeah sure Google, whatever, I don't care about Gemini models anymore. Too hard to access, constant rugpulls, intransparent pricing, many, MANY outages and in general very unreliable. At this point I just want Google to fail.
Are we sure the guy that creates google naming schemes is not held hostage and those are his desperate attempts to communicate with the outside world in concealed way?
Google team: these names are an embarrassment
The naming scheme alone has no rhyme or reason. Model names should not be a source of cognitive load.
It is no more confusing than the naming schemes used by OpenAI and Anthropic
OpenAI is at least arranged by size of celestial feature. Anthropic did arrange them by size of a written work. But I think Fable and Mythos made it a little weird. As for Google, no idea other than Hydrogen is small so at least that kind of tracks.
Having said this, I don't defend this trend at all and also think a model picker ought to be good enough with at most three tiers rather than five or so with sometimes like five effort levels each for 25 combos. I refuse to believe we need this granularity.
OpenAI - Celestial body size.
Anthropic - Literary scale.
Google - random elements with no clear relation to model value/capacity.
A fable is bigger than an opus? An astra is bigger than a sol, which is also a star?
Indeed. That's why when I had to name my product I named it after a cat.
Is this a replacement for Flash-Lite? Because it’s being used to serve answers for Search.
The new naming scheme seems shockingly poor though, “Pro” and “Flash” made intuitive sense to me, and Anthropic and OpenAI kept a sense of scale. Nobody is going to intuitively understand “Oh yes, Hydrogen is smaller than Carbon”, even though it is.
I don't believe so. From my use I think it is a Flash model through and through.
I have found it a good subagent workhorse for code implementation, with some higher level model doing the spec and review.
No. It has the same price with 3.8 flash (with no introduction price it seems)
Their super smart recursively improving ASI models didn't tell them this naming is absolutely terrible?
Are we going to talk about naming and confusing ways to access it, or are we going to discuss how capable the model is?
Every single time a Google model comes out, an army of Anthropic shills sidetrack the discussion.
Can we look at what the model does and discuss whether Google is still in the game or not? (naming models and whether you need to access it by clicking a drop-down menu or making a prayer ain't really relevant)
Another day, another nonsensically named google product.
I quit using Gemini models months ago because they are incomprehensible to understand in terms of when to use them and how to navigate the terrible GCP and Gemini UI. It’s a Kafkaesque nightmare.
With all the names for their AI offerings, I need to come up for Oxygen.
Maybe I'm not on the latest models (I do actually pay the min plan), but just this week I tried Gemini again. It still doesn't even get the current date/time of day correct. If it cant do that, I cant trust any output from it. I have no idea how Google fumbled so badly, but benchmark scores don't mean a thing if you cant even get your LLM to output the current date and time correctly.
You could vibe code a better LLM with Fable than what they have delivered. It is embarassing.
What? I just asked it what time it was in my Gemini app and it told me.
What is your prompt?
I asked about the time of market open for premarket trading for nasdaq trades. While it correctly determined when the pre markets would open, it then said it WAS open already by referencing the current time which was off by an hour. I told it where I was located and it then shifted by two hours. finally I told it to look it up and it got it correct.
Aren't Google leaders embarrassed that the first thing I think of when I see their models drop is: I will see if I can call it from openrouter?
Why are there 7 different ways to call this model from Google? Which one should I use?
Is no one at the wheel calling the shots at Google and it's a PIP every man for himself? I want to try them out, don't know how.
It's in the second paragraph of the article.
"for all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro and Ultra subscribers in Google Antigravity."
So which one of these six options is the one I should use?
Same question as whether you should be using the latest Claude Opus model from Claude Chat, Claude App, Claude Code (via Claude Pro/Max/Ultra subscription), Claude Code (via API), Claude Platform, Claude on AWS Bedrock, Claude on Azure Foundry, Claude Managed Agents, Claude Cowork, etc?
When you call it through OpenRouter you're using the API.
You know how Claude Code is like a little program that your computer can run which lets you talk to Claude? Well Gemini is actually really similar! If that's more your speed, you should use Antigravity.
But if all of this is feeling really overwhelming and you just want something simple, I would recommend just Googling it! Hydrogen will be the one generating the response :)
Whichever one suits your needs.
AI studio if you want to just try it out for some code things (think of it as a web IDE)
API if you want it to use it as an API.
For business if you are an enterprise.
Antigravity
Or just you know - just go to Google.com and search for something. It's really not hard.
This instapost "wah I don't know how to use it wah" from anyone as soon as Google release a model is getting tiring - change the record please.