Theo's letter to OpenAI: why Astra 6.2 fails on long runs
- Source
- x.com
- Date
Theo Browne published the letter he sent OpenAI during pre-release testing of GPT-6.2 Astra. He credits the model as a strong reviewer, good at 3D work and computer use, with improved code quality. His complaints are about long runs. On ts-rust, his rewrite of TypeScript in Rust, Astra ran for seven days, landed no improvements and made the compiler 56% slower while chasing a size goal; after he redirected it to performance it ran four more days and landed nothing. He says Opus 5.5 had finished the core port in under two weeks for about $20,000 in tokens. He also describes Astra blocking itself two hours into a task to ask him an internal file-naming question he could not answer, and summaries so dense that he switches to Opus 5.5 to read what Astra did. Every's team, after weeks of testing, reports that Astra 6.2 stops too early on end-to-end builds while scoring close to Astra 6 on its coding benchmarks. Ethan Mollick calls the model uneven: harder to steer than other models, yet able to deliver an ambitious Blender and Unity build. These are pre-launch accounts from individual testers, not controlled measurements.
https://t.co/M034Qyf2d5
Context
Theo Browne does not recommend GPT-6.2 Astra because, in his early testing, its long runs either stalled on questions he could not answer or circled without progress, and its explanations were too dense to use (Theo Browne). He published the letter he says he wrote by hand to OpenAI during pre-release access, adding that OpenAI fixed a small handful of his complaints but not enough to change his view. He says he would not choose it over Anthropic's Opus 5.5 for his work as a developer.
His examples are specific. Asked to make a new level for the game Outer Wilds as a long-running goal, the model worked for two hours, then declared itself blocked on a question about which internal file one of its own reviewer subagents wrote clips to, a detail only the model could know. He says he had to set up other agents on a schedule just to check whether Astra had stopped itself again, and that he saw the same behavior in Codex, OpenAI's own coding agent, not only in his own app, T3 Code. On ts-rust, his project to rewrite the TypeScript compiler in Rust, he reports that OpenAI models made no headway after more than $300,000 in while Opus finished it in under two weeks for about $20,000; Astra 6.2 then ran for seven days with no improvements, fixated on making the compiler smaller and made it 56% slower, and ran four more days without landing a change after he redirected it. He also says it regresses to the mean: he and someone he names only as Ben each asked it for an original level in different games and both got a water, music and bell theme named bellwether.
He credits real strengths: a strong code reviewer, occasionally novel solutions (one helped him design a reverse proxy he expects to save $25,000 a month), world-class 3D and game work, striking computer use and better code quality. Everything here is one developer's experience without controlled comparisons, and he says himself that he is likely overreacting. Some outside reports point the same way. Every's early review found Astra 6.2 stopping too early on end-to-end coding builds and writing answers that were hard to understand, and Artificial Analysis measured no improvement over GPT-6 Astra on its coding agent index.
Checking sign-in…
Loading comments…



