Jev's speed numbers hold up, Theo says not for compaction
Guillermo Rauch's team measured Jev up to 18x faster at p95 than GPT Luna on a safety classifier, and more accurate. The same week, Theo called using Jev to compact agent history a terrible strategy, which is the argument worth reading.
The big disagreement
Theo calls Jev compaction terrible, Armin Ronacher will not dismiss it
Theo listed six problems with scoring each tool call and deleting what a 32k context model judges irrelevant: it cannot see encrypted reasoning traces, and deleting early history forces cache rewrites that cost more than leaving the text in. Armin Ronacher replied that he would not dismiss it, because most harnesses need pruning on compaction anyway for cost reasons.
This is a terrible compaction strategy that fundamentally doesn't understand how compaction and context management work.
Receipts
Jev's speed claims survive two independent classifier benchmarks
Two separate teams ran TypeSafe's Jev against the small models they were already using for classification, and it won on both speed and quality.
Jev is up to 18x faster (p95) *and* more accurate.
Three hours, four model passes, and ThePrimeagen hated the result
ThePrimeagen built a feature with Fable, had Sol review it, grok do a simplicity pass and Sol check the simplifications, and says it produced one of the worst interfaces he has ever seen. David K reported a related failure on his own code: Fable 5.1 wrongly assumed an extra URL redirect and then deleted the E2E tests for that case, which he says would have broken prod.
and all of this to produce one of the worst interfaces I have ever seen
How they actually work
Boris Cherny stopped managing sessions after Claude Code Projects
Projects landed in Claude Code as one conversation that splits work into threads itself, runs them as parallel cloud sessions and keeps going after you close your laptop, in beta for select users. Boris Cherny says he now just sends thoughts as they come and the project remembers how he works.
I stopped managing sessions.
Model and agent watch
Open weight models hit 78% of tokens on Vercel as teams cap the expensive ones
Guillermo Rauch says open models reached 78.4% of token volume on Vercel AI Gateway, with Moonshot AI and DeepSeek at third and fourth by spend and those plus Z.ai together passing OpenAI. Gergely Orosz talked to an AI-pilled company that went from unlimited budgets to a daily limit on Astra and Fable, with Sol and Opus still unlimited.
We now have a daily budget limit for SOTA models
Frequently asked questions
What is Jev compaction and why did Theo criticize it?
Jev compaction is a technique that scores each tool call and deletes what a 32k context model judges irrelevant. Theo listed six problems with it, including that it cannot see encrypted reasoning traces and that deleting early history forces cache rewrites that cost more than leaving the text in.
Did Jev perform well in independent testing?
Yes, two separate teams ran Jev against the small models they were already using for classification, and it won on both speed and quality. The results showed interesting latency improvements inside agent loops for decide, route and score calls.
What went wrong when ThePrimeagen used Fable with multiple AI reviewers?
ThePrimeagen built a feature with Fable and had it reviewed and simplified by multiple agents, but the result was one of the worst interfaces he had ever seen. Review agents check code at face value, so bugs from unstated assumptions can survive because only reading the actual diff catches them.
What are Claude Code Projects and how do they work?
Claude Code Projects is a feature that keeps one conversation that splits work into threads automatically, runs them as parallel cloud sessions, and continues after you close your laptop. The project remembers how you work so you can send thoughts as they come instead of managing sessions yourself.
Are companies still using expensive frontier AI models or switching to cheaper options?
Open weight models now represent 78.4% of token volume on Vercel AI Gateway. Teams are increasingly adopting daily spending limits on expensive models while routing tasks to cheaper options instead of always reaching for the frontier model.