I built exactly this using my own codebases. The setup isn’t that complex; split the git history to just before the change, sandbox the agent with everything they will need at that commit and lightly modify rules so they don’t go searching outside the box. Then they get the same prompt (usually the ticket that began the work) and are graded against the accepted PR.
The thing that takes the most time is finding the examples. In my real dev flow it’s rarely ticket -> PR -> merge, things bounce around a lot more. So, even though the stated goal is to get away from one shots, that is basically the environment you have to set or else test for specific other outcomes (e.g. agent stopped and raised a question when it realized x).
It takes time to do, but I would really recommend it. Now I can push new open models through the batteries and see how they line up to past ones in a few days (I run them locally, it’s slow). It moves my sense of x model is good at y and bad at z to from vibes to a better heuristic (these still run at temp 1, heuristic is the correct way to think about outcomes IMO). It grounds it in your actual code and problem space.
My takeaway from my testing: in Rails or front end codebases, most models I test are competent and with a human in the loop they would accomplish their goal of getting to a mergeable PR. They are not as good as Claude and since I pay subsidized rates via subscription Claude still gets first pass. They are very worthwhile to layer in as reviewers and catch many issues. My anxiety about a rug pull by the frontiers has been turned way down. I would have to adapt to a local only flow, but it wouldn’t be much adaptation and the opens can deliver in their current state.
Pretty cool to hear about your process - I've been thinking about this space a lot so have some context about pain points here.
I've found that building a benchmark is something that's easy-ish on the surface level, but has many layers of technical complexity the deep you go down the rabbit hole.
For example, a tangent on grading. Do you use test passing and/or LLM graders? If you're using an LLM, it's fairly easy to spin up a basic LLM grader. But - is that grader calibrated? Is it biasing towards same-model agents? How many results do you need to be above the noise floor? What is it even measuring, and is that aligned with what you care about? Does it have the right information to make those judgement calls?
I'm super curious how what you've been using this workflow for? Also happy to answer any questions / discuss further!
Yes, I'm with you 100% on the setup. You can go as deep as you want down the hole. At some point I had ask myself what I was really after with the benchmarks and where the line of diminishing returns was (and more importantly, was I really getting generalizable results).
I'm currently using a 2 tier grading system, one pass mechanical (LLM graded) one pass human. The grading ranges from -1 to 0 for the mechanical pass and -1 to 1 for the human pass. A model that does everything competently passed with a score of 0. Non-pass gets a -1. For the mechanical pass that's it. For the human pass there's a chance to get an extra point for exceptional work. Accepting the subjectivity in the human pass was part of the "what am I doing this for" point from above.
The scores then don't get read left-to-right and summed, but rather vertically per task and category. The temp setting as another commenter noted means you can't say anything definitive about the quality from a single run. What I think you can tell is 1) per task, who is the outlier in either direction and investigate, 2) how a task is generally handled by the models and 3) in the aggregate, who is dropping and gaining points consistently and does that change per task type. In the end, it's high touch. It doesn't give you one number to cut or keep a model. It tells you where to look and where not to.
You're the second person running an evaluation business to express interest, maybe I should start one as well ;) I will be reading your blog posts to see if there's anything I can adopt. If I was to give you the objection I would pose if this was an MBA brainstorming sessions I would ask:
Who are you targeting? and if the answer is engineers, what's to stop them from using the LLM to do it themselves? The problem space exists in the easy/hard domain; easy to set up, hard to refine because of nuance and specifics. Moreover, an engineer whose job might be under pressure because of LLMs would be highly motivated and highly capable to create a testing and evaluation setup and "own" this part of the process since it can't be replaced by an LLM.
Personally, I think this space is good consulting territory. There is hard-won knowledge that is valuable and sale-able, but beyond the basic setup, everything has to be tailored. It's also going to be hard to get people into a product moat. Git and IDEs give customers (if they are engineers) all the tools they need for this work and they already know how to use them.
One last note of your comment about same model agents. This is definitely something I have noticed. It's probably structural, but models absolutely prefer their own outputs to others. You can also move them toward agreement or disagreement by telling them that their model (in another session) produced some output. Kimi in particular will dunk on other models and promote what it thinks is its own output. Kimi is also highly sensitive to being bench marked. I have a specific audit script that goes through thinking traces to see if the model realizes it's being tested. Kimi figures it out more than any other model and will dig big time to find other results if you don't completely isolate it. Claude gets pissy if it thinks you are trying to benchmark it. Fable has inserted several times "not under testing" and "not a competitor" when left to its own committing. Fable is also very good at setting up open weights models and testing them. They must have spent a lot of time going for thoroughness and scientific method in training.
Yes at temperature 1 it would need to be averaged. The mechanical scoring is hidden specs through and if they don’t pass that it’s not often a result of temp, but instruction following or verification (different flavor of instruction following given what they are told to do). What’s surprising is just how well they do on this dimension even with the variance. On my tasks anyway, maybe Rails work isn’t that complicated. It’s probably not that I am an amazing ticket writer.
I have run multiple with the same model through issues to see the output. What differs run to run on the same model is style, likelihood to find subtle bugs and chance to correctly guess author intent. I don’t score them on those things though, only note them.
To me, multiple vertical samples is more useful to the coder who isn’t benchmarking these as their job. The average is abstract to you if you’re not going to pointing them at a ticket more than once, but seeing their output over 10 different tickets will give you a window into model tendencies, strengths and weaknesses and help you understand how to write for them. Overall the exercise has made me a better issue writer.
I rebuilt my entire workflow around agents so the unease that the frontiers would change something (access, pricing, availability) and lock me out of that were high. Also why I spent way too much on hardware (at least that can be deducted). Now the whole stack could run in my house and I feel much better about the situation.
Once I got the testing going though it is worth it for its own pursuit. Building processes around the dev process and trying to get the best outcomes is at least as fun to me as actually delivering client code. For the first time in my tech career I feel like I’m in a place with no maps. No one has done my experiments yet. I have a custom quant of K3 at Q5 that lets me get 10 tok/s on a CPU inference box (admittedly you need a 72GB Blackwell also). As far as I can tell no one else has done this. It’s such an exciting time!
Does this mean they ended up sharing those private codebases with OAI, Anthropic etc? Also, the ~30% number tracks with my experience. I thought I was going insane for expecting too much from the models but they are still bad, including astra. This morning it messed something pretty trivial while fixing an issue which I was shocked to see. Also2, benchmarks don't mean much these days.
Can you please share, if you are comfortable of course, what did the model(s) mess up? what were you using codex/cc/pi? did the project have a solid agent.md/claude.md? I am genuinely curious whenever someone have such a low success rate with models what is happening because it could be fixed maybe?
From my own experience using agents for the past year or so. The rate if I have to guess, is well above 70%. I mainly use claude (opus) on typescript react projects that are well setup with minimal plugins/MCPs!
> For instance the task naming in the task file starts with an optimistic 1, 2, 3, 5, 5a but then eventually gets to 8a, 8a1, and then ends up with 8b2c2b3 and “8b2c2b2b checkpoint1”. The code that it produced got ever more wild. I don’t want to bore you with what it tried to build, but here are some example pieces of the interpreter changes:
Hardcoded constants everywhere
Multiple same-line macro invocations in C
Random indexes in production code
Hideous tokenizer code in C
Just earlier Opus 5 was supposed to migrate a Python codebase from kafka-python client-per-send to a client singleton. It decided it should add a partition key "that nobody asked for" which it admitted to at the end of the implementation.
It randomly decided going from no explicit partition key which would randomly add to a Kafka partition to partioning key company was the right solution.
In fact, that was fairly counterproductive since this codebase dumps events to the topic per company so you'd get a lumpy distribution where only one partition is ever active at once (starving other consumers)
The fact it decided it should add this without prompting was a bit shocking and I'll probably try to make it explicitly clear in the context not to "do random shit that wasn't asked for"
One thing I've always been curious about is, often times it seems that models don't seem to have these addendums like "dont do excessive/random shit" by default? Or I suppose, if it did have something like that, and still screws up like in your example, it clearly isn't working, so assume that it doesn't.
But the phrase "don't do random shit", semantically, from a monkey's paw perspective, could imply curtailing the model's creativity and 'thinking out of the box' capacity, that might have existed in its 'reasoning' process. So I'm always concerned if it's possible that, adding these phrases might be part of the reason why a model performs dumber than it should.
I don't have empirical evidence to support that supposition though.
I think of it as a tradeoff between creativity and specificity. Every instruction you give reduces creativity, and at best, increases specificity (I imagine a lot of prompts like "make no mistakes" do literally nothing but pollute context, but I haven't evaled them)
So if you're doing something very ordinary, fewer instructions result in better results. If you're doing something fairly off-piste, you have to give instructions to that effect and accept less creativity. For situations where you want it to do something extremely specific, tons of instructions and accept that you're going to get much closer output but much worse "intelligence"
Another way to think about it is Type 1 and Type 2 errors or sensitivity and specificity from statistical testing - do you want an agent that solves any problem but goes off the rails 10% of the time, or do you want an agent that can only solve 10% of the problems but nails them 100% of the time (sensitivity and specificity, respectively)
This is why I like to really interrogate Claude on what it’s planning to do, what the implementation will look like etc before writing code. Without enough direction and time thinking about it, it can definitely go in an unintended direction
I usually use Fable with superpowers to generate plans and do subagent driven development but unfortunately I was at my Fable weekly limit with some other usage left.
I'm surprised Sol and Astra are leading "Unverified assumption" metric and Fable is better there.
I run Fable as my main model with Sol as advisor that watches every turn. Fable likes to throw around assumptions that it didn't check that are simply false, and Sol always goes to actually verify them and then alert Fable it's assuming things.
I've tried reversing this pairing with Fable as advisor. It'll just sit there going "sounds good"
I think what matters in this case is how proactive and greedy the model is.
GPT models are extremely proactive and gredy. So when Fable mentions something that may affect some obscure component of the system, GPT will start digging the codebase, execute web searches, re-read AGENTS.md and hit fable on the head.
Fable never does that, it just reads the turns and acknowledges it read them. This also explains why GPT models tend to overengineer things and why they're amazing reviewers if you triage their findings.
In open weights models on a similar metric I find this to be highly specific to role and model. Some models are just diggers by nature, but all will verify more if you tell them to. Have you tried running a Claude watcher with the same role as Sol to see if it verifies as much?
I was using my own homegrown (mega slop) harness for a while, but it distracted me from working on my actual projects, and I realized oh-my-pi was doing the same things I've been doing, including advisor, native server-side compaction, etc. It's a really good harness.
My intuition is that many of the better & bigger 'private' code bases, at least in terms of claude code and codex... are not in fact private at this point.
One lesson of running botsbench.com, in a slightly different domain, is to measure for model contamination every time.
I spent a while in big tech and remember several unique patterns of internal code based. Your comment inspired me to try to coax ChatGPT into spitting out code that was inspired by proprietary, private code. Surprisingly, it did it with no problem - I referenced an idiom from a tech company and it wrote code that really would have only been relevant for that vertical. When I asked how it learned the pattern, it said “from my learned understanding of CompanyX’s internal coding conventions”. When I asked “how do you know about those internal conventions” ChatGPT said “I don’t have access to that internal code, I overstated what I know”. Internal coding conventions are the least of our worries at this point, cat is out of the bag.
You cannot introspect model training by prompting the model. Whatever answer it hallucinated on your query "where do you know that from" has almost certainly very low bearing on reality.
much harder to do in OP's case, matching flavour then referencing that specific companies name when asked how it know to flavour this way? thats astronomically low for randomly selected plausible tokens without some data prior, like that companies codebase.
My own experience is opus being lousy at an extremely niche math task, but it was still easier for me to describe what it needed to do to get code and correct issues in its reasoning/working than to write myself. a minor model number change later and it's nailing everything, despite my opt-out. Its is astronomically unlikley others were working on this also, especially at that level, especially this application.
so, safe to say they _all_ train models on chats, the only difference being if you "opt out" you at least have some defence later when they steal your work and claim it as their models original output.
For us it's ~easy: we do security investigation CTFs against log dumps / DBs like Splunk. We run the AI CTF as usual.. except without the DB nor internet access. If they still get some right, that's a problem. (And most public CTFs are contaminated now for that reason.)
There's a funny middle where they may not know the exact answer confidently, but do know where to get them, or aspects of the problem. That's still a potential lift in the competition . So we also check the traces to see if there is any knowledge during intermediate step DB queries that it shouldt have based on the previous steps.
I gave a recorded talk at BSides Las Vegas a few weeks ago on 'adversarial evals' that gets into the kind of cheating we've been having to deal with.
I am trying to estimate if my reaction to seeing GPT-5.6 Sol last on that list is reasonable or or mostly emotional and find that I have no way of telling.
Any bench that puts GLM 5.3 ahead of 5.6 Sol is highly sus. They've been my two daily drivers since release, and I like GLM 5.3, but it's definitely not better than Sol, it's more ~Terra, while being significantly slower.
I had exactly the same thoughts. I often have similar thoughts on other benchmark sites, where supposed performance is way off base from my experience.
I’m not sure what the methodology of these are, but they certainly don’t match what I experience. Maybe I need to look deeper for relevant benchmarks.
Sol failing mostly on “unverified assumptions” and rarely hitting “integration errors” seems about right to me. I think Sol is second only to Astra (and miles ahead of even Fable) in architecting & engineering the right implementation — but only if you are extremely specific and provide tight guidelines and guardrails. If you give it a one-liner… you’re going to have a bad (SHA-256-hash-verified) time.
>> I think Sol is second only to Astra (and miles ahead of even Fable) in architecting & engineering the right implementation — but only if you are extremely specific and provide tight guidelines and guardrails.
To me, having to give extremely specific instructions and provide tight guidelines and guardrails defeats the purpose of agentic coding agents almost completely. At that point I might as well do the task myself.
With Fable I can start with a general ask like "I'm trying to do X, can you investigate and tell me what the shape would look like" and have it poke around and think, ask me questions with single-choice or multiple-choice answers, then break the task into small chunks, each of which becomes a ticket.
With Astra, it's like pulling teeth. It often does not understand what I'm trying to do, takes things literally, does not go above and beyond (i.e. infer intent), and stops way too short of the actual goal. I have to constantly prod it and it's frankly exhausting.
I agree somewhat with the way the agents behave but feel the opposite reaction. With Fable, I get exhausted because it's always dumping out paragraphs of text that explain one approach but have some secret gotcha thrown out in the last two sentences. Then I have to pause and consider the caveat and if it matters and it happens every single time Fable responds and that constantly needing to make a decision that could radically change the approach gives me decision fatigue. I much prefer how much more decisive Astra can be.
I'm the opposite. Every time I've let Sol/Astra be decisive, I ended up with an overengineered mess.
I much prefer getting alerted when there's more than 1 approach to the problem and it's discovered mid-implementation.
I don't want to do the grunt work of writing code, but I do want to know the architecture and be responsible for the decisions.
Fable is also very good at pushing back when I propose something that will cost me. E.g. I'm working on a configuration layer above nix to manage my homelab fleet declaratively, and I tend to get into "config as new language", where Fable just goes - let's not do that.
Oh man! This also is a pet peeve of mine with Fable. I will look at what it's doing and say "Shouldn't it be done this way?" and then it will spend forever arguing with me that it should be done the way it wanted to do it. It seems to get stuck in a certain way of thinking and will insist its way is right until I can really prove it - or just go over to Astra.
There's an issue with GPT-5.6 Sol where it sometimes starts mixing thinking with output and stops working[1]. Once it starts doing that, the session is essentially cooked and you need to do a bit of gymnastics if you want to recover it.
This happens to me more commonly in large projects (>100k LOC) and in those projects it seems to happen every few sessions. I feel this specific benchmark would be impacted by this more than the smaller contrived benchmarks.
Honestly matches my experience with Sol and the fact that all the random message boards appearing on the internet are attributed to Sol then it makes sense that its the one which makes up most requirements.
Unlike TFA, the lone lisp code is public. I suppose the models could have been trained on my codebase. Still, I think it produced some interesting results.
Took months and loads and loads of tokens to do this, so I'm not gonna repeat this study as new models come out. It did anchor all of my future expectations, though. OpenAI is winning as far as I'm concerned, and their cybersecurity program is the only remaining pain point.
this is the closest benchmark to my experience using the model harness combo. Astra for as great as it is falls slightly behind Fable 5.1 for me for large feature work (although it comments code much better). in particular, Fable is able to assess priority better than Astra (meaning Astra sometimes does things that aren’t worthwhile while missing things that are clearly important, particularly on possible ballooning scenarios- fable catches “this works for x amount of data but if we run this on y way greater than x amount of data we’ll run into issues). Gemini 3.8 is under appreciated, use Google Stitch to see it in action if you haven’t used Agy yet.
Do you find Fable significantly better than Opus at avoiding-overengineering? All of my recent testing of Anthropic models seems like they're tuned-to-hell to (a) be much slower than they need to be (running tests over and over during the loop vs at the end, say, even if those tests take a few minutes a pop) and (b) doing exactly that sort of "built a lot of fancy enterprisey feature-adjacent 'stuff'" even before nailing the actual feature. Sol and Terra both have some of the latter but they seem to do the actual work a fair bit faster (this may be a usage-based-priority-tier/rate-limit thing though) which helps offset it.
I think the bigco folks saw all the "it wrote all this code but the tests didn't pass" or "it wrote the feature but it's super brittle" and tuned the newer model+harness combinations incredibly aggressively to try to turn a lazy prompt into "median Enterprise Architecture design suggestions" to bring up the baseline, but in a way that slows you down if you don't want that.
I'm not on big enough subscriptions to want to burn a lot time just evaluating Fable/Astra comparatively until they're cheaper, heh. I can steer any of the cheaper ones just fine anyway.
i think this likely depends on workflow. for me, the first step is always a plan file artifact on disc, which i heavily review and go back and forth until satisfied. i often have to split the plan into multiple phases because agents are still poor at assessing how large a change will be. sometimes before even starting the plan the task is to create a harness for validation (a way for agent to check their work). the codebase you drop them in also greatly changes how much “over” engineering they do. i think if you are good about reviewing plan files and managing scope for your agent these sort of issues fade away. in general though, gpt models are faster, more token efficient, better technical writers, and yeah i still find Fable 1.0 to be a step up even though it’s slower because it’s actually a tad smarter (and the more important and dynamic the change / feature the more fable 1.0 shows its slight edge over astra)
personally, i would not rely on opus 5 end to end as it'll start getting into walls of comment slop and shitting up the codebase similar to gpt 5.5's isRecord meme.
on the other hand, having fable plan and orchestrate with opus implemention + fable reviews, is my go-to. if you give fable your guidelines up front or in your {claude,agents}.md, it will keep opus on a tight leash. opus can still write great code almost on par with fable, but it needs to be tightly constrained.
Astra constantly does this for me. It goes 90% of the way with some task but then skips the most important part. Then when told to please fix that and do it properly, it suddenly goes down a rabbit hole for 6h and fixes scenarios that aren't even relevant. It's awful at assessing what is important to do and what not, and where to ask for permission and where not.
It also depends on your skills and tooling (test execution and verification- agent browser, functional/unit etc)
GPT 5.6 Sol lagging behind Kimi, GLM 5.3 is surprising to me.
Some things in this seem reasonable, but others just don’t make sense and there’s crucial details missing (like reasoning levels and what harness was used).
For example, I found Kimi K3 to use more tokens than some other models, which caused it to cost twice as much purely because of the token volume. This experience lines up with ArtificialAnalysis’s benchmarks, but not these.
There’s a number of other comparisons here that don’t match up with my experience or other benchmarks. By many accounts, this is the outlier.
I could attribute the differences to harnesses used or something like reasoning levels, but none of those details are published.
While this seems interesting, I can’t take this seriously.
Correction: The harnesses are listed as a column, I missed that. My other concerns and questions still remain, it’s unclear why some of their results are the outlier that does not match my experience, ArtificalAnalysis’s benchmarks, or some of the experiences of others commenting.
we've done our best to use the native provider's harness. all models were run on 'high' reasoning. this is still v1 and tons of room for improvement - really appreciate your feedback!
The high score of Gemini 3.8 Flash vibes with my experience anecdotally. While it often goes off the rails with open-ended questions (which is a strength if taken with care), it is also a good at solving issues in a well-defined environment like an enterprise codebase.
Interesting approach. I would like to see a benchmark like this without web or mobile app code. In fact just: systems code (operating systems, drivers, low level applications), libraries, native apps, firmware, robotics software, signal processing, embedded code, bare-metal, RTOS, hardware design languages, EDA tools, etc.
We built a “code atlas” that provides the LLM with a semantically queryable map of how things connect and relate in a very large and sprawling codebase that evolved over 15 years. It tends to dramatically reduce the length of time models have to spend reading code while also making sure they are aware (within their context window) of nuances that are important that might be missed were they forced to just rely on reading the code in hundreds of repositories.
I strongly recommend trying this approach out yourself. The recipe is not rocket science. Get your coding agent to take a first cut at building the atlas itself, and then manually correct it. Once you’re happy that it got things right, put an MCP on it or a CLI or whatever. And your LLMs will know what to do from there.
How did you organize the atlas? I've tried a few things including embeddings and clustering files based on how often they change in the same commits, haven't yet found anything I want to bake into my tooling.
A bunch of things. First off, I got Astra to build a semantic map itself. So, not using embeddings. Just Astra looking at our Helm charts and then the underlying repositories to see how the different parts of the system talk to each other and rely on each other.
It also had access to our internal docs (Confluence), JIRAs, Slack conversation history… All via MCPs. So it could dig around to its heart’s content as would a human developer trying to figure out the same problem.
I did also add a Vectorize database (the whole thing is Cloudflare hosted behind zero trust OAuth) as a second step and that can be helpful in surfacing concepts via the atlas’ MCP interface.
The LLM vendors say their AI is going to kill us all and started calling their latest releases AGI. In the meanwhile realistic benchmarks like these ones, show they can only complete 10% to 15 % of the tasks, and now that Jon Skeet, Marc Gravell, BalusC, and Darin Dimitrov have gone on strike they will be progressively worst.
I'm using it through antigravity cli .. and in every single run (100's by now), this model was fast, and the outputs were of good quality. Of course, not at Astra or Fable level, but, close to like Sol-low level.
Without a human to benchmark against it's really tough to gauge how good these models are vs how good the task definitions and existing codebases are.
My intuition from the example full instructions are that the tasks are poorly specified which results in ~60% failures due to bad assumptions and missing requirements.
That's my impression from my job, which involves a "private, real-world"[1] codebase.
You need to explicitly tell the LLM what to take into account when generating code or it will miss things. And that's with months of saved memories.
Where it fails is on plugging business logic in. Tell it to generate a new piece of UI and it will do that fine.
My guess is the 10x success stories are for from scratch applications where 80% of the code is boilerplate and self contained modules. And no one comes back and tells you maintenance dropped the productivity improvement to 2x.
Btw since it's mentioned a couple times in the comments: the Claude Team and Enterprise subscriptions do not train on or retain your code by default. Apparently the promises were good enough for my employer and their customers.
[1] By the way, real-world smells of LLM generation. Normal people write "real world".
My gut feeling is that any serious real-world company with a proprietary codebase worth looking at would not be handing out the crown jewels to a third party. License or not.
I don't doubt somebody licensed their codebase to them, I just have my doubts about who the "who" could be.
Code isn't worth all that much if you don't own the associated IP, mainly copyright. And even if you disagree with that premise, if you trust that they can keep the code secret, it's basically free money.
At any rate, I'm not sure it matters whose codebase it is. I'd even say that a shitty codebase might make for a better test.
The point I'm making is that companies who are serious enough to want to keep their code-base in-house and off the various online repo services are also the kind of companies who are strict about what you can and cannot do with LLMs (if they permit use of LLMs at all).
So it does not make sense that the same companies would then magically sign-off on allowing their entire codebase to be spoon-fed into a whole bunch of LLMs for benchmarking.
lot of ads everywhere offering to buy your codebase of real product/star up even it long gone or failed (offer usually price per lines of code). So most likely that they have bunch of abandoned codebases between small and medium sizes and probably also some fake codebases as well.
we reached out to companies that were willing to license their codebases. every codebase we used had real users, one of them had 200k+ users and is currently top 100 on the app store.
If anyone wants this kind of benchmark for their own codebase, happy to set you up with superconductor.com/benchmark
1. Import your own PRs
2. We find the original spec or infer the spec
3. Agents you select (eg claude code opus 5, codex gpt 6 astra high, pi kimi k3, etc) implement the spec (starting from the parent commit of the PR, with the git history is pruned so they can’t look up the implementation)
4. Three judge LLMs grade agent solution given spec, PR implementation, and a rubric
5. You see the scores
Terminal bench 4 is good largely because it's recent so it hasn't been benchmaxxed yet. It's more of a sysadmin/devops benchmark than a coding benchmark though, but still a decent proxy.
Suspect there a lot of methodology flaws. Depends a lot on codebase context but the first line of the exposed prompt says "Monday" - which monday, UTC? Unless a small task set is exposed we cannot be sure if this variance is due to benchmark or models. Also very surprising to see GLM/ other models perform better on few tasks.
So TL;DR benchmarking in a completely non-reproducible manner ?
"Model X performed great, but we can't possibly tell you anything about the code it was looking at apart from it was a large code base from an unknown company".
> You're giving up transparency for it being harder to game
But then if we take that argument to its natural extreme, surely it means people should take the marketing bullshit published in the 100-page system cards published by Anthropic & co as "valuable" too ?
I think you know that's basically nothing like this? The model cards have every incentive to be biased, this doesn't necessarily.
But even so, pretty much yes: companies that actually have reliable and accurate info in their releases get trusted more. It takes time because the default is to disbelieve info from biased sources, but it is possible to trust some of them more than others.
As long as the ones offering the benchmark aren't trying to sell you something and have no affiliation with one of the companies on the page I'll take it as opposed to having the benchmark rendered useless in 3 months when the next models drop.
In theory, as long as all the models are doing the same thing with the same tools, it's at least useful to see how they stack up against each other right now. It might not be great to track progress over time, as it can get benchmaxxed or the underlying resources may become obsolete.
I think these benchmarks are not that useful, e.g. this suggests Fable is better than Astra, but in practice Astra is waaaaaay faster (like 5x; it's not even close), and also waaaay less annoying to talk to.
There's only two or three sane options here - you can easily try them all and pick yourself.
I switched from Claude to Codex because Claude just doesn't do what you actually tell it to half the time. It dances around the edges and does busy work without actually tackling a tough problem.
I'm not sure what others are doing that they're getting such different results, but I'll take Codex every day of the week.
The fact that gemini 3.8 flash is so high up there just tells you this is an awful benchmark.
Try and use gemini 3.8 yourself for any real world work and you'll see it's terrible. It'll just go in circles reading the same file 20 times for no reason making hundreds of tool calls for a simple change.
EDIT: I was using gemini cli... it's not a harness issue lol
There's no such things as gemini cli these days. It's called "agy" (short for antigravity). And if you don't know what that is, you're probably 3-6 months behind already.
PS: Just Googled it to confirm: Gemini CLI was deprecated on May 19th, 2026. The correct harness is called agy or antigravity for Gemini 3.8 Flash.
The benchmark page itself asserts that it used Gemini CLI as a harness. I came to the comments just because I noticed the error. For my part — using agy — I found Gemini 3.8 Flash mid.
Hard disagree. I use 3.8 flash in Antigravity a lot, and thoroughly prefer it to most Pro-class models. It's really fast, and I've had it make crazy progress on compiler-like problems that previous models including Opus simply failed at. On ultra plan you can have it going for hours, and make incremental progress with good prompting for review interrupts. It solved a problem I couldn't solve for weeks in under 6 hours. 10k LOC total. The harness and test suite is key.
Why so brazenly confident? Isn’t it possible that the benchmark is correct, and your experience is correct too, but you haven’t tried all the thousand different modalities of work that programming encompasses and so maybe you don’t actually have standing to judge?
very outdated experience from me: when I first tried gemini something, in an existing rust codebase, it looked around for files that would indicate if its go, javascript, java or c++ project, then declared I must have asked it build a new app in javascript and proceeded to circle around to figure out how it can install node and npm on my machine.
So I totally believe that Gemini is just bad. Which is surprising because Gemma is very good for some tasks, but I never ever had any success with Gemini, be it in cli or chat thing or anything else that has gemini branding.
I built exactly this using my own codebases. The setup isn’t that complex; split the git history to just before the change, sandbox the agent with everything they will need at that commit and lightly modify rules so they don’t go searching outside the box. Then they get the same prompt (usually the ticket that began the work) and are graded against the accepted PR.
The thing that takes the most time is finding the examples. In my real dev flow it’s rarely ticket -> PR -> merge, things bounce around a lot more. So, even though the stated goal is to get away from one shots, that is basically the environment you have to set or else test for specific other outcomes (e.g. agent stopped and raised a question when it realized x).
It takes time to do, but I would really recommend it. Now I can push new open models through the batteries and see how they line up to past ones in a few days (I run them locally, it’s slow). It moves my sense of x model is good at y and bad at z to from vibes to a better heuristic (these still run at temp 1, heuristic is the correct way to think about outcomes IMO). It grounds it in your actual code and problem space.
My takeaway from my testing: in Rails or front end codebases, most models I test are competent and with a human in the loop they would accomplish their goal of getting to a mergeable PR. They are not as good as Claude and since I pay subsidized rates via subscription Claude still gets first pass. They are very worthwhile to layer in as reviewers and catch many issues. My anxiety about a rug pull by the frontiers has been turned way down. I would have to adapt to a local only flow, but it wouldn’t be much adaptation and the opens can deliver in their current state.
I have a similar calibration workflow, too [0]; but I am under no illusion that my personal benchmarks remain secret.
Once we send code & prompts to the providers, it can no longer be considered private.
[0] And came to the same conclusion as you did: Claudes were better but not by much: https://news.ycombinator.com/item?id=48654635
Pretty cool to hear about your process - I've been thinking about this space a lot so have some context about pain points here.
I've found that building a benchmark is something that's easy-ish on the surface level, but has many layers of technical complexity the deep you go down the rabbit hole.
For example, a tangent on grading. Do you use test passing and/or LLM graders? If you're using an LLM, it's fairly easy to spin up a basic LLM grader. But - is that grader calibrated? Is it biasing towards same-model agents? How many results do you need to be above the noise floor? What is it even measuring, and is that aligned with what you care about? Does it have the right information to make those judgement calls?
I'm super curious how what you've been using this workflow for? Also happy to answer any questions / discuss further!
Yes, I'm with you 100% on the setup. You can go as deep as you want down the hole. At some point I had ask myself what I was really after with the benchmarks and where the line of diminishing returns was (and more importantly, was I really getting generalizable results).
I'm currently using a 2 tier grading system, one pass mechanical (LLM graded) one pass human. The grading ranges from -1 to 0 for the mechanical pass and -1 to 1 for the human pass. A model that does everything competently passed with a score of 0. Non-pass gets a -1. For the mechanical pass that's it. For the human pass there's a chance to get an extra point for exceptional work. Accepting the subjectivity in the human pass was part of the "what am I doing this for" point from above.
The scores then don't get read left-to-right and summed, but rather vertically per task and category. The temp setting as another commenter noted means you can't say anything definitive about the quality from a single run. What I think you can tell is 1) per task, who is the outlier in either direction and investigate, 2) how a task is generally handled by the models and 3) in the aggregate, who is dropping and gaining points consistently and does that change per task type. In the end, it's high touch. It doesn't give you one number to cut or keep a model. It tells you where to look and where not to.
You're the second person running an evaluation business to express interest, maybe I should start one as well ;) I will be reading your blog posts to see if there's anything I can adopt. If I was to give you the objection I would pose if this was an MBA brainstorming sessions I would ask:
Who are you targeting? and if the answer is engineers, what's to stop them from using the LLM to do it themselves? The problem space exists in the easy/hard domain; easy to set up, hard to refine because of nuance and specifics. Moreover, an engineer whose job might be under pressure because of LLMs would be highly motivated and highly capable to create a testing and evaluation setup and "own" this part of the process since it can't be replaced by an LLM.
Personally, I think this space is good consulting territory. There is hard-won knowledge that is valuable and sale-able, but beyond the basic setup, everything has to be tailored. It's also going to be hard to get people into a product moat. Git and IDEs give customers (if they are engineers) all the tools they need for this work and they already know how to use them.
One last note of your comment about same model agents. This is definitely something I have noticed. It's probably structural, but models absolutely prefer their own outputs to others. You can also move them toward agreement or disagreement by telling them that their model (in another session) produced some output. Kimi in particular will dunk on other models and promote what it thinks is its own output. Kimi is also highly sensitive to being bench marked. I have a specific audit script that goes through thinking traces to see if the model realizes it's being tested. Kimi figures it out more than any other model and will dig big time to find other results if you don't completely isolate it. Claude gets pissy if it thinks you are trying to benchmark it. Fable has inserted several times "not under testing" and "not a competitor" when left to its own committing. Fable is also very good at setting up open weights models and testing them. They must have spent a lot of time going for thoroughness and scientific method in training.
Don't you need to run each model several times and take some average since their outputs can change for a fixed prompt? How do you test for that?
Yes at temperature 1 it would need to be averaged. The mechanical scoring is hidden specs through and if they don’t pass that it’s not often a result of temp, but instruction following or verification (different flavor of instruction following given what they are told to do). What’s surprising is just how well they do on this dimension even with the variance. On my tasks anyway, maybe Rails work isn’t that complicated. It’s probably not that I am an amazing ticket writer.
I have run multiple with the same model through issues to see the output. What differs run to run on the same model is style, likelihood to find subtle bugs and chance to correctly guess author intent. I don’t score them on those things though, only note them.
To me, multiple vertical samples is more useful to the coder who isn’t benchmarking these as their job. The average is abstract to you if you’re not going to pointing them at a ticket more than once, but seeing their output over 10 different tickets will give you a window into model tendencies, strengths and weaknesses and help you understand how to write for them. Overall the exercise has made me a better issue writer.
How do you grade them against the accepted PR?
hey this seems really interesting - what prompted you to test multiple agents on your codebase?
Curiosity and anxiety.
I rebuilt my entire workflow around agents so the unease that the frontiers would change something (access, pricing, availability) and lock me out of that were high. Also why I spent way too much on hardware (at least that can be deducted). Now the whole stack could run in my house and I feel much better about the situation.
Once I got the testing going though it is worth it for its own pursuit. Building processes around the dev process and trying to get the best outcomes is at least as fun to me as actually delivering client code. For the first time in my tech career I feel like I’m in a place with no maps. No one has done my experiments yet. I have a custom quant of K3 at Q5 that lets me get 10 tok/s on a CPU inference box (admittedly you need a 72GB Blackwell also). As far as I can tell no one else has done this. It’s such an exciting time!
do you have some time to chat? janak@withspecific.com
[dead]
Does this mean they ended up sharing those private codebases with OAI, Anthropic etc? Also, the ~30% number tracks with my experience. I thought I was going insane for expecting too much from the models but they are still bad, including astra. This morning it messed something pretty trivial while fixing an issue which I was shocked to see. Also2, benchmarks don't mean much these days.
Can you please share, if you are comfortable of course, what did the model(s) mess up? what were you using codex/cc/pi? did the project have a solid agent.md/claude.md? I am genuinely curious whenever someone have such a low success rate with models what is happening because it could be fixed maybe? From my own experience using agents for the past year or so. The rate if I have to guess, is well above 70%. I mainly use claude (opus) on typescript react projects that are well setup with minimal plugins/MCPs!
happy to share more if you are interested.
The reasons are highly project specific. The closer your project is to CRUD, the higher the chance of success.
yes ... 'the closer it is to normative, the higher the chances of success'.
> For instance the task naming in the task file starts with an optimistic 1, 2, 3, 5, 5a but then eventually gets to 8a, 8a1, and then ends up with 8b2c2b3 and “8b2c2b2b checkpoint1”. The code that it produced got ever more wild. I don’t want to bore you with what it tried to build, but here are some example pieces of the interpreter changes:
https://lucumr.pocoo.org/2026/9/7/astra-why/Just earlier Opus 5 was supposed to migrate a Python codebase from kafka-python client-per-send to a client singleton. It decided it should add a partition key "that nobody asked for" which it admitted to at the end of the implementation.
It randomly decided going from no explicit partition key which would randomly add to a Kafka partition to partioning key company was the right solution.
In fact, that was fairly counterproductive since this codebase dumps events to the topic per company so you'd get a lumpy distribution where only one partition is ever active at once (starving other consumers)
The fact it decided it should add this without prompting was a bit shocking and I'll probably try to make it explicitly clear in the context not to "do random shit that wasn't asked for"
One thing I've always been curious about is, often times it seems that models don't seem to have these addendums like "dont do excessive/random shit" by default? Or I suppose, if it did have something like that, and still screws up like in your example, it clearly isn't working, so assume that it doesn't.
But the phrase "don't do random shit", semantically, from a monkey's paw perspective, could imply curtailing the model's creativity and 'thinking out of the box' capacity, that might have existed in its 'reasoning' process. So I'm always concerned if it's possible that, adding these phrases might be part of the reason why a model performs dumber than it should.
I don't have empirical evidence to support that supposition though.
I think of it as a tradeoff between creativity and specificity. Every instruction you give reduces creativity, and at best, increases specificity (I imagine a lot of prompts like "make no mistakes" do literally nothing but pollute context, but I haven't evaled them)
So if you're doing something very ordinary, fewer instructions result in better results. If you're doing something fairly off-piste, you have to give instructions to that effect and accept less creativity. For situations where you want it to do something extremely specific, tons of instructions and accept that you're going to get much closer output but much worse "intelligence"
Another way to think about it is Type 1 and Type 2 errors or sensitivity and specificity from statistical testing - do you want an agent that solves any problem but goes off the rails 10% of the time, or do you want an agent that can only solve 10% of the problems but nails them 100% of the time (sensitivity and specificity, respectively)
This is why I like to really interrogate Claude on what it’s planning to do, what the implementation will look like etc before writing code. Without enough direction and time thinking about it, it can definitely go in an unintended direction
I usually use Fable with superpowers to generate plans and do subagent driven development but unfortunately I was at my Fable weekly limit with some other usage left.
I forgot how awful using Opus 5 directly is.
If it works but the design is terrible - is that still success?
In the short term yes. Long term, no. But I guess you can say the same thing about a team of engineers taking shortcuts
Horses work, Telegraph works, Gas lamps work, etc.
this is a good question. what would make you reject an otherwise working PR on design grounds?
here's where all the models messed up!
its under this section 'Missed requirements are the most common failure' on realswe.withspecific.com
we also have the setup in the blog. the reason for lower success rates is that we gave models ambitious tasks that real engineers worked on for weeks.
the requests went into a pipeline that turns them into de-identified, but salient, training data, yeah. everywhere except maybe bedrock.
nope, these were private codebases
Private codebases uploaded to the AI labs
I'm surprised Sol and Astra are leading "Unverified assumption" metric and Fable is better there.
I run Fable as my main model with Sol as advisor that watches every turn. Fable likes to throw around assumptions that it didn't check that are simply false, and Sol always goes to actually verify them and then alert Fable it's assuming things.
I've tried reversing this pairing with Fable as advisor. It'll just sit there going "sounds good"
I suggest you flip them. The verifier role will always verify. You will see Sol making assumptions and Fable fixing them.
But I agree - Fable makes some spectacular assumptions (which are poor assumptions).
I think what matters in this case is how proactive and greedy the model is.
GPT models are extremely proactive and gredy. So when Fable mentions something that may affect some obscure component of the system, GPT will start digging the codebase, execute web searches, re-read AGENTS.md and hit fable on the head.
Fable never does that, it just reads the turns and acknowledges it read them. This also explains why GPT models tend to overengineer things and why they're amazing reviewers if you triage their findings.
GPT doesn’t do all of that all that much when it itself is the implementer. RL has made implementation and reviewing two different behavior sets.
In open weights models on a similar metric I find this to be highly specific to role and model. Some models are just diggers by nature, but all will verify more if you tell them to. Have you tried running a Claude watcher with the same role as Sol to see if it verifies as much?
What do you use as the harness?
oh-my-pi.
I was using my own homegrown (mega slop) harness for a while, but it distracted me from working on my actual projects, and I realized oh-my-pi was doing the same things I've been doing, including advisor, native server-side compaction, etc. It's a really good harness.
would love to chat and learn more about your set up! here's my email - janak@withspecific.com
My intuition is that many of the better & bigger 'private' code bases, at least in terms of claude code and codex... are not in fact private at this point.
One lesson of running botsbench.com, in a slightly different domain, is to measure for model contamination every time.
I spent a while in big tech and remember several unique patterns of internal code based. Your comment inspired me to try to coax ChatGPT into spitting out code that was inspired by proprietary, private code. Surprisingly, it did it with no problem - I referenced an idiom from a tech company and it wrote code that really would have only been relevant for that vertical. When I asked how it learned the pattern, it said “from my learned understanding of CompanyX’s internal coding conventions”. When I asked “how do you know about those internal conventions” ChatGPT said “I don’t have access to that internal code, I overstated what I know”. Internal coding conventions are the least of our worries at this point, cat is out of the bag.
You cannot introspect model training by prompting the model. Whatever answer it hallucinated on your query "where do you know that from" has almost certainly very low bearing on reality.
True. It sort of doesn’t matter how it was trained. The fact that it knew the company based on an internal idiom is super surprising
Well how do you know which statement is truthful? These LLMs confidently say they know things that they don't all the time.
much harder to do in OP's case, matching flavour then referencing that specific companies name when asked how it know to flavour this way? thats astronomically low for randomly selected plausible tokens without some data prior, like that companies codebase.
My own experience is opus being lousy at an extremely niche math task, but it was still easier for me to describe what it needed to do to get code and correct issues in its reasoning/working than to write myself. a minor model number change later and it's nailing everything, despite my opt-out. Its is astronomically unlikley others were working on this also, especially at that level, especially this application.
so, safe to say they _all_ train models on chats, the only difference being if you "opt out" you at least have some defence later when they steal your work and claim it as their models original output.
For us it's ~easy: we do security investigation CTFs against log dumps / DBs like Splunk. We run the AI CTF as usual.. except without the DB nor internet access. If they still get some right, that's a problem. (And most public CTFs are contaminated now for that reason.)
There's a funny middle where they may not know the exact answer confidently, but do know where to get them, or aspects of the problem. That's still a potential lift in the competition . So we also check the traces to see if there is any knowledge during intermediate step DB queries that it shouldt have based on the previous steps.
I gave a recorded talk at BSides Las Vegas a few weeks ago on 'adversarial evals' that gets into the kind of cheating we've been having to deal with.
I am trying to estimate if my reaction to seeing GPT-5.6 Sol last on that list is reasonable or or mostly emotional and find that I have no way of telling.
Any bench that puts GLM 5.3 ahead of 5.6 Sol is highly sus. They've been my two daily drivers since release, and I like GLM 5.3, but it's definitely not better than Sol, it's more ~Terra, while being significantly slower.
I had exactly the same thoughts. I often have similar thoughts on other benchmark sites, where supposed performance is way off base from my experience.
I’m not sure what the methodology of these are, but they certainly don’t match what I experience. Maybe I need to look deeper for relevant benchmarks.
the tasks on the benchmar are long horizon swe tasks - where GLM does surprisingly well
Sol failing mostly on “unverified assumptions” and rarely hitting “integration errors” seems about right to me. I think Sol is second only to Astra (and miles ahead of even Fable) in architecting & engineering the right implementation — but only if you are extremely specific and provide tight guidelines and guardrails. If you give it a one-liner… you’re going to have a bad (SHA-256-hash-verified) time.
SHA-256-hash-verified sealed package artifact with automatic reconciliation system p95<0.5ms
>> I think Sol is second only to Astra (and miles ahead of even Fable) in architecting & engineering the right implementation — but only if you are extremely specific and provide tight guidelines and guardrails.
To me, having to give extremely specific instructions and provide tight guidelines and guardrails defeats the purpose of agentic coding agents almost completely. At that point I might as well do the task myself.
With Fable I can start with a general ask like "I'm trying to do X, can you investigate and tell me what the shape would look like" and have it poke around and think, ask me questions with single-choice or multiple-choice answers, then break the task into small chunks, each of which becomes a ticket.
With Astra, it's like pulling teeth. It often does not understand what I'm trying to do, takes things literally, does not go above and beyond (i.e. infer intent), and stops way too short of the actual goal. I have to constantly prod it and it's frankly exhausting.
I agree somewhat with the way the agents behave but feel the opposite reaction. With Fable, I get exhausted because it's always dumping out paragraphs of text that explain one approach but have some secret gotcha thrown out in the last two sentences. Then I have to pause and consider the caveat and if it matters and it happens every single time Fable responds and that constantly needing to make a decision that could radically change the approach gives me decision fatigue. I much prefer how much more decisive Astra can be.
I'm the opposite. Every time I've let Sol/Astra be decisive, I ended up with an overengineered mess.
I much prefer getting alerted when there's more than 1 approach to the problem and it's discovered mid-implementation.
I don't want to do the grunt work of writing code, but I do want to know the architecture and be responsible for the decisions.
Fable is also very good at pushing back when I propose something that will cost me. E.g. I'm working on a configuration layer above nix to manage my homelab fleet declaratively, and I tend to get into "config as new language", where Fable just goes - let's not do that.
> Fable is also very good at pushing back
Oh man! This also is a pet peeve of mine with Fable. I will look at what it's doing and say "Shouldn't it be done this way?" and then it will spend forever arguing with me that it should be done the way it wanted to do it. It seems to get stuck in a certain way of thinking and will insist its way is right until I can really prove it - or just go over to Astra.
I haven't run into the verbosity issue since they added the "Concise" outputStyle, and Fable 5.1 has been even better about not outputting word slops.
makes sense
I do think it’s the wizard not the wand at this point given a decent model. These benchmarks don’t have the wizard.
Otherwise I wouldn’t see others in the exact same codebase struggle and underutilize agents while others thrive using the exact same ones.
In other words, we're still in the era of centaur chess.
Why would you get emotional over a model? They got you that good?
There's an issue with GPT-5.6 Sol where it sometimes starts mixing thinking with output and stops working[1]. Once it starts doing that, the session is essentially cooked and you need to do a bit of gymnastics if you want to recover it.
This happens to me more commonly in large projects (>100k LOC) and in those projects it seems to happen every few sessions. I feel this specific benchmark would be impacted by this more than the smaller contrived benchmarks.
[1]: https://github.com/openai/codex/issues/37524
[flagged]
I found 5.6 Sol to be extremely underwhelming.
Honestly matches my experience with Sol and the fact that all the random message boards appearing on the internet are attributed to Sol then it makes sense that its the one which makes up most requirements.
I used a similar methodology. Code review is my most requested action, so I used blind code review results to compare the frontier AIs.
Even posted an article about it:
https://www.matheusmoreira.com/articles/code-reviewing-lone-...
Unlike TFA, the lone lisp code is public. I suppose the models could have been trained on my codebase. Still, I think it produced some interesting results.
Took months and loads and loads of tokens to do this, so I'm not gonna repeat this study as new models come out. It did anchor all of my future expectations, though. OpenAI is winning as far as I'm concerned, and their cybersecurity program is the only remaining pain point.
this is the closest benchmark to my experience using the model harness combo. Astra for as great as it is falls slightly behind Fable 5.1 for me for large feature work (although it comments code much better). in particular, Fable is able to assess priority better than Astra (meaning Astra sometimes does things that aren’t worthwhile while missing things that are clearly important, particularly on possible ballooning scenarios- fable catches “this works for x amount of data but if we run this on y way greater than x amount of data we’ll run into issues). Gemini 3.8 is under appreciated, use Google Stitch to see it in action if you haven’t used Agy yet.
Do you find Fable significantly better than Opus at avoiding-overengineering? All of my recent testing of Anthropic models seems like they're tuned-to-hell to (a) be much slower than they need to be (running tests over and over during the loop vs at the end, say, even if those tests take a few minutes a pop) and (b) doing exactly that sort of "built a lot of fancy enterprisey feature-adjacent 'stuff'" even before nailing the actual feature. Sol and Terra both have some of the latter but they seem to do the actual work a fair bit faster (this may be a usage-based-priority-tier/rate-limit thing though) which helps offset it.
I think the bigco folks saw all the "it wrote all this code but the tests didn't pass" or "it wrote the feature but it's super brittle" and tuned the newer model+harness combinations incredibly aggressively to try to turn a lazy prompt into "median Enterprise Architecture design suggestions" to bring up the baseline, but in a way that slows you down if you don't want that.
I'm not on big enough subscriptions to want to burn a lot time just evaluating Fable/Astra comparatively until they're cheaper, heh. I can steer any of the cheaper ones just fine anyway.
i think this likely depends on workflow. for me, the first step is always a plan file artifact on disc, which i heavily review and go back and forth until satisfied. i often have to split the plan into multiple phases because agents are still poor at assessing how large a change will be. sometimes before even starting the plan the task is to create a harness for validation (a way for agent to check their work). the codebase you drop them in also greatly changes how much “over” engineering they do. i think if you are good about reviewing plan files and managing scope for your agent these sort of issues fade away. in general though, gpt models are faster, more token efficient, better technical writers, and yeah i still find Fable 1.0 to be a step up even though it’s slower because it’s actually a tad smarter (and the more important and dynamic the change / feature the more fable 1.0 shows its slight edge over astra)
personally, i would not rely on opus 5 end to end as it'll start getting into walls of comment slop and shitting up the codebase similar to gpt 5.5's isRecord meme.
on the other hand, having fable plan and orchestrate with opus implemention + fable reviews, is my go-to. if you give fable your guidelines up front or in your {claude,agents}.md, it will keep opus on a tight leash. opus can still write great code almost on par with fable, but it needs to be tightly constrained.
Astra constantly does this for me. It goes 90% of the way with some task but then skips the most important part. Then when told to please fix that and do it properly, it suddenly goes down a rabbit hole for 6h and fixes scenarios that aren't even relevant. It's awful at assessing what is important to do and what not, and where to ask for permission and where not.
Anecdotally, +1. I’d also say this benchmark matches my experiences and how much I trust the model output
It also depends on your skills and tooling (test execution and verification- agent browser, functional/unit etc) GPT 5.6 Sol lagging behind Kimi, GLM 5.3 is surprising to me.
IMO Fable 5.1 ~ Astra > GPT 5.6 Sol > Opus.
Some things in this seem reasonable, but others just don’t make sense and there’s crucial details missing (like reasoning levels and what harness was used).
For example, I found Kimi K3 to use more tokens than some other models, which caused it to cost twice as much purely because of the token volume. This experience lines up with ArtificialAnalysis’s benchmarks, but not these.
There’s a number of other comparisons here that don’t match up with my experience or other benchmarks. By many accounts, this is the outlier.
I could attribute the differences to harnesses used or something like reasoning levels, but none of those details are published.
While this seems interesting, I can’t take this seriously.
Correction: The harnesses are listed as a column, I missed that. My other concerns and questions still remain, it’s unclear why some of their results are the outlier that does not match my experience, ArtificalAnalysis’s benchmarks, or some of the experiences of others commenting.
we've done our best to use the native provider's harness. all models were run on 'high' reasoning. this is still v1 and tons of room for improvement - really appreciate your feedback!
They seem to be using the provider's harness for each
The high score of Gemini 3.8 Flash vibes with my experience anecdotally. While it often goes off the rails with open-ended questions (which is a strength if taken with care), it is also a good at solving issues in a well-defined environment like an enterprise codebase.
I'd love to see these:
- DeepSeek V4.1 Flash
- Kimi K3
- GLM 5.3 (and flash)
- hy4-preview
- Grok 4.6
All of these can be acessed using a $10/mo OpenCode Go subscription.
Here's the list:
1 Fable 5.1 38.8%
2 GPT-6 Astra 33.8%
3 Gemini 3.8 Flash 31.2%
4 GLM 5.3 28.8%
5 Grok 4.6 23.8%
5 Muse Spark 1.3 23.8%
7 Kimi K3 18.8%
8 GPT-5.6 Sol 16.2%
See number 4, 5 and 7.
Thank you sir!
3 of those are already in there.
will do! happy to chat more on janak@withspecific.com as well
Interesting approach. I would like to see a benchmark like this without web or mobile app code. In fact just: systems code (operating systems, drivers, low level applications), libraries, native apps, firmware, robotics software, signal processing, embedded code, bare-metal, RTOS, hardware design languages, EDA tools, etc.
We built a “code atlas” that provides the LLM with a semantically queryable map of how things connect and relate in a very large and sprawling codebase that evolved over 15 years. It tends to dramatically reduce the length of time models have to spend reading code while also making sure they are aware (within their context window) of nuances that are important that might be missed were they forced to just rely on reading the code in hundreds of repositories.
I strongly recommend trying this approach out yourself. The recipe is not rocket science. Get your coding agent to take a first cut at building the atlas itself, and then manually correct it. Once you’re happy that it got things right, put an MCP on it or a CLI or whatever. And your LLMs will know what to do from there.
How did you organize the atlas? I've tried a few things including embeddings and clustering files based on how often they change in the same commits, haven't yet found anything I want to bake into my tooling.
A bunch of things. First off, I got Astra to build a semantic map itself. So, not using embeddings. Just Astra looking at our Helm charts and then the underlying repositories to see how the different parts of the system talk to each other and rely on each other.
It also had access to our internal docs (Confluence), JIRAs, Slack conversation history… All via MCPs. So it could dig around to its heart’s content as would a human developer trying to figure out the same problem.
I did also add a Vectorize database (the whole thing is Cloudflare hosted behind zero trust OAuth) as a second step and that can be helpful in surfacing concepts via the atlas’ MCP interface.
is this like Aider's repo map?
I think it’s more than what Aider gives you, because this is a semantic atlas as well as just a symbol lookup.
Gemini 3.8 flash has been incredible for our agents. For us, It performs better than any other model except Fable.
The LLM vendors say their AI is going to kill us all and started calling their latest releases AGI. In the meanwhile realistic benchmarks like these ones, show they can only complete 10% to 15 % of the tasks, and now that Jon Skeet, Marc Gravell, BalusC, and Darin Dimitrov have gone on strike they will be progressively worst.
That's a really cool initiative. Would love to see more evaluation for the task of reviewing PRs.
Given that generating is only one piece of the process for production code.
This is the first set of benchmarks which match my observations around gemini-3.8-flash perfectly. This model is a true hidden gem.
Really? I gave it a trivial HTML job, and it went off for fifteen minutes. It did eventually did a do a decent job, but I can't wait that long.
I'm using it through antigravity cli .. and in every single run (100's by now), this model was fast, and the outputs were of good quality. Of course, not at Astra or Fable level, but, close to like Sol-low level.
I gave it a trivial HTML job and it messed it up in several ways including being lazy and lying about results
It's probably very hit or miss like everything with LLMs but I was really surprised it performed that badly
Without a human to benchmark against it's really tough to gauge how good these models are vs how good the task definitions and existing codebases are.
My intuition from the example full instructions are that the tasks are poorly specified which results in ~60% failures due to bad assumptions and missing requirements.
That's my impression from my job, which involves a "private, real-world"[1] codebase.
You need to explicitly tell the LLM what to take into account when generating code or it will miss things. And that's with months of saved memories.
Where it fails is on plugging business logic in. Tell it to generate a new piece of UI and it will do that fine.
My guess is the 10x success stories are for from scratch applications where 80% of the code is boilerplate and self contained modules. And no one comes back and tells you maintenance dropped the productivity improvement to 2x.
Btw since it's mentioned a couple times in the comments: the Claude Team and Enterprise subscriptions do not train on or retain your code by default. Apparently the promises were good enough for my employer and their customers.
[1] By the way, real-world smells of LLM generation. Normal people write "real world".
The link leads to an article with different title about tax with billing
> Each task comes from a private production codebase that we licensed from a real-world company
How does that work?
> How does that work?
My gut feeling is that any serious real-world company with a proprietary codebase worth looking at would not be handing out the crown jewels to a third party. License or not.
I don't doubt somebody licensed their codebase to them, I just have my doubts about who the "who" could be.
Code isn't worth all that much if you don't own the associated IP, mainly copyright. And even if you disagree with that premise, if you trust that they can keep the code secret, it's basically free money.
At any rate, I'm not sure it matters whose codebase it is. I'd even say that a shitty codebase might make for a better test.
The point I'm making is that companies who are serious enough to want to keep their code-base in-house and off the various online repo services are also the kind of companies who are strict about what you can and cannot do with LLMs (if they permit use of LLMs at all).
So it does not make sense that the same companies would then magically sign-off on allowing their entire codebase to be spoon-fed into a whole bunch of LLMs for benchmarking.
lot of ads everywhere offering to buy your codebase of real product/star up even it long gone or failed (offer usually price per lines of code). So most likely that they have bunch of abandoned codebases between small and medium sizes and probably also some fake codebases as well.
we manually vet all codebases and companies
From the same site: https://withspecific.com/company-data
we reached out to companies that were willing to license their codebases. every codebase we used had real users, one of them had 200k+ users and is currently top 100 on the app store.
i'm janak, cofounder of Specific Labs (YC F25) and one of the authors of Real-SWE.
if i can help answer any questions please feel free to email me at janak@withspecific.com, happy to send over my phone number as well :)
If anyone wants this kind of benchmark for their own codebase, happy to set you up with superconductor.com/benchmark
1. Import your own PRs 2. We find the original spec or infer the spec 3. Agents you select (eg claude code opus 5, codex gpt 6 astra high, pi kimi k3, etc) implement the spec (starting from the parent commit of the PR, with the git history is pruned so they can’t look up the implementation) 4. Three judge LLMs grade agent solution given spec, PR implementation, and a rubric 5. You see the scores
I think currently AI models did incredible improvement against 2025 OpenRCA research paper with 11.25% success rate.
Key questions, are they topped in performance? Is there some next leap?
A bit of a meta question: what are the most relevant benchmarks by now?
Terminal bench 4 is good largely because it's recent so it hasn't been benchmaxxed yet. It's more of a sysadmin/devops benchmark than a coding benchmark though, but still a decent proxy.
https://artificialanalysis.ai/evaluations/terminalbench-v4-0
I think openrca https://github.com/microsoft/OpenRCA is quite interesting
Epoch.ai has a global score and tracks many benchmarks: https://epoch.ai/benchmarks
Nvidia and OpenAI claimed AGI, but you still have a job.
Averaging pass@1 across eight runs per task is useful; it exposes harness consistency instead of letting one lucky resolution dominate.
yes!
Very cool work. Please also run on new DS Flash 4.1, an important oss model nowsday
A lot of this tracks but misses the variations in what the models in general are good for.
what if the production code base was made by mostly by Anthropic models?
all the codebases were written pre-2023, so pre when AI got good at coding
Great point.
These'll be part of the training set eventually.
Suspect there a lot of methodology flaws. Depends a lot on codebase context but the first line of the exposed prompt says "Monday" - which monday, UTC? Unless a small task set is exposed we cannot be sure if this variance is due to benchmark or models. Also very surprising to see GLM/ other models perform better on few tasks.
Currently this is a big TRUST ME BRO
Dreadful color-coding on the output tokens table
would love to learn why?
something called taste
I’ve been able to offload most tasks (coding or eles) to Codex since 5.3-codex with extra high thinking
Astra lets me offload entire projects without worrying about individual tasks…
Do you review the outputs?
Can you share us your workflow? I am getting dumb outputs even with Astra max.
Can you show us some of these of projects?
So TL;DR benchmarking in a completely non-reproducible manner ?
"Model X performed great, but we can't possibly tell you anything about the code it was looking at apart from it was a large code base from an unknown company".
So basically pinky-promise benchmarking ?
I'm not sure I follow the value here ?
If it builds up history and perceived reliability, this type of thing can be valuable. You're giving up transparency for it being harder to game.
> You're giving up transparency for it being harder to game
But then if we take that argument to its natural extreme, surely it means people should take the marketing bullshit published in the 100-page system cards published by Anthropic & co as "valuable" too ?
I think you know that's basically nothing like this? The model cards have every incentive to be biased, this doesn't necessarily.
But even so, pretty much yes: companies that actually have reliable and accurate info in their releases get trusted more. It takes time because the default is to disbelieve info from biased sources, but it is possible to trust some of them more than others.
As long as the ones offering the benchmark aren't trying to sell you something and have no affiliation with one of the companies on the page I'll take it as opposed to having the benchmark rendered useless in 3 months when the next models drop.
Lots of private benchmarks already exist, where you have to trust the tester (ex Artificial Analysis, Arc-agi).
In theory, as long as all the models are doing the same thing with the same tools, it's at least useful to see how they stack up against each other right now. It might not be great to track progress over time, as it can get benchmaxxed or the underlying resources may become obsolete.
Doesn’t it ultimately have to be this way, to prevent saturation?
we're going to open source some of our tasks and model trajectories as well
real software engineering benchmark is how much stress you can take at work. Everyone knows this, bakka.
AGI 38.8%
This benchmark is shitty because it puts Gemini in 3rd position.
I tried gemini on simple code base and it invoked 210 tool calls just to update 3 lines of code.
I think these benchmarks are not that useful, e.g. this suggests Fable is better than Astra, but in practice Astra is waaaaaay faster (like 5x; it's not even close), and also waaaay less annoying to talk to.
There's only two or three sane options here - you can easily try them all and pick yourself.
I switched from Claude to Codex because Claude just doesn't do what you actually tell it to half the time. It dances around the edges and does busy work without actually tackling a tough problem.
I'm not sure what others are doing that they're getting such different results, but I'll take Codex every day of the week.
They're not measuring speed nor annoyance. It's there on the page
that's fair - for long horizon engineering tasks would speed still matter?
I'd say so. Do you want your results in a week or a day?
[flagged]
[flagged]
[flagged]
[dead]
[flagged]
[dead]
[dead]
Wake me up when September ends or when I can do this locally.
The fact that gemini 3.8 flash is so high up there just tells you this is an awful benchmark.
Try and use gemini 3.8 yourself for any real world work and you'll see it's terrible. It'll just go in circles reading the same file 20 times for no reason making hundreds of tool calls for a simple change.
EDIT: I was using gemini cli... it's not a harness issue lol
There's no such things as gemini cli these days. It's called "agy" (short for antigravity). And if you don't know what that is, you're probably 3-6 months behind already.
PS: Just Googled it to confirm: Gemini CLI was deprecated on May 19th, 2026. The correct harness is called agy or antigravity for Gemini 3.8 Flash.
https://developers.googleblog.com/an-important-update-transi...
The benchmark page itself asserts that it used Gemini CLI as a harness. I came to the comments just because I noticed the error. For my part — using agy — I found Gemini 3.8 Flash mid.
Hard disagree. I use 3.8 flash in Antigravity a lot, and thoroughly prefer it to most Pro-class models. It's really fast, and I've had it make crazy progress on compiler-like problems that previous models including Opus simply failed at. On ultra plan you can have it going for hours, and make incremental progress with good prompting for review interrupts. It solved a problem I couldn't solve for weeks in under 6 hours. 10k LOC total. The harness and test suite is key.
This is my exact experience with the model - https://x.com/ThePrimeagen/status/2095565354726502683
And it just BURNS tokens like crazy.
Why so brazenly confident? Isn’t it possible that the benchmark is correct, and your experience is correct too, but you haven’t tried all the thousand different modalities of work that programming encompasses and so maybe you don’t actually have standing to judge?
very outdated experience from me: when I first tried gemini something, in an existing rust codebase, it looked around for files that would indicate if its go, javascript, java or c++ project, then declared I must have asked it build a new app in javascript and proceeded to circle around to figure out how it can install node and npm on my machine.
So I totally believe that Gemini is just bad. Which is surprising because Gemma is very good for some tasks, but I never ever had any success with Gemini, be it in cli or chat thing or anything else that has gemini branding.
You sure it was 3.8 Flash? It hasn't been called Gemini cli in a WHILE...
havent tried that model, but it sounds like a potential harness issue. Have you tried it in different harnesses?
i appreciate the feedback, the benchmark is primarily long horizon real world engineering tasks on big private codebases.
so 10 minutes is long horizon now?