Rubi ships a token X-ray that reads Claude Code's local transcripts. On 30 days of real usage:
Finding (284 sessions, 30 days)
Share of cost that was re-reading context (cache reads) 71%
Share of cost from sessions that grew past 150k tokens 97%
[...]"
Now that last number, above, is interesting... 97% of the cost originating from sessions where the context window is > 150k.
It's almost like there is, or could be, two AI/LLM marketplaces -- one for sessions with context windows < 150k -- a "cheap" or "cheaper" or "less expensive" (heck "inexpensive" if you prefer, tier), and a more expensive, more computational resources hungry, > 150k tokens in the session "expensive" tier.
In fact, we could almost consider decision models like Jev (and open source alternatives to it that run on minimum consumer hardware) that usually respond to a single prompt and return a very short response, or tool-calling models that generate short tool calls in response to what are usually single prompts, as the least-expensive tier of LLM's, followed by regular LLM's with contexts < 150k, followed by regular LLM's with contexts > 150k...
In other words, there seems to be a hierarchy there, existing or emerging...
Anyway, Rubi looks like an interesting way to potentially save a lot on Claude conversations!
>"1.3 What the data showed
Rubi ships a token X-ray that reads Claude Code's local transcripts. On 30 days of real usage:
Finding (284 sessions, 30 days)
Share of cost that was re-reading context (cache reads) 71%
Share of cost from sessions that grew past 150k tokens 97%
[...]"
Now that last number, above, is interesting... 97% of the cost originating from sessions where the context window is > 150k.
It's almost like there is, or could be, two AI/LLM marketplaces -- one for sessions with context windows < 150k -- a "cheap" or "cheaper" or "less expensive" (heck "inexpensive" if you prefer, tier), and a more expensive, more computational resources hungry, > 150k tokens in the session "expensive" tier.
In fact, we could almost consider decision models like Jev (and open source alternatives to it that run on minimum consumer hardware) that usually respond to a single prompt and return a very short response, or tool-calling models that generate short tool calls in response to what are usually single prompts, as the least-expensive tier of LLM's, followed by regular LLM's with contexts < 150k, followed by regular LLM's with contexts > 150k...
In other words, there seems to be a hierarchy there, existing or emerging...
Anyway, Rubi looks like an interesting way to potentially save a lot on Claude conversations!