GPT-6.2 Astra Launch Draws Coordinated 'Don't Use' From Prominent AI Figures
GPT-6.2 Astra released, but its reputation has collapsed...
Today in the early morning, GPT released GPT-6.2 Astra. A strange phenomenon: I didn't see a single post about the model release on my timeline homepage. I wonder if it's because the model has been criticized too much...
This time GPT-6.2 Astra and Sol were released together, and the official release page still looks quite cool. 
I originally wanted to write a review, but by the time I saw it, it was too late. I missed the first wave of hype, so I gave up. As for the review content, you can find detailed descriptions from many official accounts.
One after another.
After this GPT-6.2 release, I saw three well-known AI big names jointly issue a statement:
Do not recommend using the GPT-6.2 Astra model
Matt's post is extremely simple: one sentence, not recommended.
Theo, even though he's sick in bed, wrote a long summary and emphasized that it was completely hand-typed without AI polish, just like what you see from me in this article.

Let me talk about his main points.
Theo said he doesn't like GPT-6.2 Astra, and he will continue to use Opus 5.5.
First, the pros:
- Top-tier performance in 3D and game development.
- Computer use is ridiculously strong.
- Code quality noticeably improved.
- Good at writing subagent prompts.
Apart from those pros, everything else is cons. There are too many problems, so I'll pick a few key ones:
- GPT-6.2 Astra has no concept of knowledge blind spots: when doing long-horizon tasks, it can use a large number of subagents, and it manages them fairly well, but it asks some strange, absurd, mindless questions, and these questions are unrelated to the coding agent.
- As usual, it doesn't speak human language. Its explanations are only understandable to itself, simply not meant for humans. Its descriptions only see the trees, not the forest, getting trapped in details. And it will even assert that the forest only has trees, which is hard to swallow. No matter how you adjust the prompts, this situation doesn't improve.
- It easily falls into infinite loops, and the kind that ignore prompts.
Dan Shipper thinks:

Roughly: writing has regressed again, and it speaks even less human. Performance is very similar to Astra 6, but when coding, Astra 6.2 often stops directly, and AI Slop is still severe.
And I also tested GPT-6.2 Astra with several demos. Taking the volcano example I previously liked, this time the effect is like this....
![[层状火山-360度环绕.mp4]]
I also made a rain curtain greenhouse demo, like this.
![[yumuwenshi.mov]]
How the results are, judge for yourselves.
These are the evaluations of the first batch of testers on long-horizon tasks.
If you're using it for cases and demos, I think you won't get these conclusions (although the demos I tested above don't show obvious progress).
And these conclusions can only be verified when integrated into your real workflow.
So don't use demos to claim that GPT-6.2 Astra has improved and the price has dropped; that's actually useless. If demos don't improve, unless OpenAI stops playing.
It can't be that the more you train, the worse it gets, right? And especially after a delayed release.
There are also big V's who think these first testers might have been paid, which I didn't expect.
I also don't believe Theo is someone who would lie. Ever since he posted an article criticizing Google, I know this guy says what he wants to say and is quite straightforward.
In short, whether a large model is good or not is not determined by whose voice is loudest.
There's a judging criterion: within the next 15 days, will you still want to use this model, and can you remember it, and once a new model is released, will you still be willing to use this model?
OpenAI, please put more effort in.