Theo calls GPT-6.2 Astra the worst model OpenAI has shipped
Two separate prompts for a creative game level came back with the same name, and OpenAI's Tibo says the early testers used an older build.
Theo says GPT-6.2 Astra is the worst model OpenAI has ever released, posted while he was, by his own account, sick in bed with a 102 degree fever and promising more detail later.
The detail he spoiled early is the strange part. During early testing, he and Ben each asked Astra for a creative new level for a game they love, Theo for Outer Wilds and Ben for Majora's Mask. They prompted separately. Both got water levels with music themes built around bells, and both levels were named "Bellwether". Theo's read is that the model collapses toward the average answer harder than anything he has used, and he asked other people to try the same prompt and report whether they also land on Bellwether.
This model regresses to the mean like nothing I've ever seen.
OpenAI says the reviewers tested an old build
@thsottiaux, who posts as Tibo, replied that part of the gap between influencer reactions and general reaction is timing. A few people got access a while ago, then a lot of changes landed in the last three or four days, and he is fairly sure nobody re-tested before the videos went out. He also said the models are a big jump over their GPT predecessors and that the teams want future iterations to improve on taste and niceness of use.
Theo took that seriously enough to retest. His verdict after a few more hours is that things are better than when he first tested but still, in his words, directionally very bad.
The test case is a slow worktree
He keeps using one question. While debugging performance running T3 Code on a VPS, two of three machines were fast and one was slow. The real cause was that the slow box uses a virtual disk while the others have physical NVMe drives. Asked outright why it was slow, pre-fix Astra produced an unnecessary graph, bolded irrelevant text, and talked about pnpm, file systems and RAM about as much as drive speed. Opus, given the same question, opened in bold with slow disk made worse by VDO. Post-fix Astra was much more concise but still, Theo says, weighted toward second-order causes.
That is the complaint in one line. It is not that the model is wrong, it is that it does not rank. Theo quotes Michael Seibel on aiming for 100% clarity and 80% accuracy rather than 100% accuracy and 50% clarity, and says he wishes OpenAI models would lie to him a little more so he could get the vibe without wading through details.
For working engineers, that is a property benchmarks do not score, and it is exactly the property that decides whether a model is useful when you are staring at a slow machine and want the cause, not an inventory.
It doesn't miss anything. It's SO thorough.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.
