2026.11.19

An article called "Reasons for Disliking GPT-6.2 Astra"

There are three main points. ① GPT-6.2 Astra shows high capability in code review, 3D and game development, Computer Use, and more. On the other hand, in long autonomous work, it apparently has a problem of stopping the work with questions about internal implementation that the user can't answer. ② After it runs a complex task, its explanation leans toward technical details, so it's hard to understand what it actually did in the end. The author puts into practice a split between models: implementing with Astra and summarizing the results with Opus 5.5. ③ Even if an AI is highly capable, that doesn't necessarily mean it delivers results in real work. In the author's testing, there was even a case where running it for 7 days brought no improvement, and the compiler instead became 56% slower. In long autonomous runs, checking progress and verifying the results are essential. For now, it seems good to delegate implementation and code review to it (for implementation, decide in advance what to build). Also, the tendency to be bad at grasping the big picture and to make local optimizations seems likely to be there in Astra 6.2 too, so it seems better to have Opus handle overall progress management.

An article called “Reasons for Disliking GPT-6.2 Astra”

There are three main points!

① GPT-6.2 Astra shows high capability in code review, 3D and game development, Computer Use, and more. On the other hand, in long autonomous work, it apparently has a problem of stopping the work with questions about internal implementation that the user can’t answer.

② After it runs a complex task, its explanation leans toward technical details, so it’s hard to understand what it actually did in the end. The author puts into practice a split between models: implementing with Astra and summarizing the results with Opus 5.5.

③ Even if an AI is highly capable, that doesn’t necessarily mean it delivers results in real work. In the author’s testing, there was even a case where running it for 7 days brought no improvement, and the compiler instead became 56% slower. In long autonomous runs, checking progress and verifying the results are essential.

For now, it seems good to delegate implementation and code review to it (for implementation, decide in advance what to build)

Also, the tendency to be bad at grasping the big picture and to make local optimizations seems likely to be there in Astra 6.2 too, so it seems better to have Opus handle overall progress management.

https://x.com/suna_gaku/status/2122827310672027872