比特币橙子Trader|11月 18, 2026 03:06
Holy crap, GPT-6.2 might really have been released just to boost Anthropic's stock price; there's a lot of negative feedback from English-speaking users.
Theo (@theo, founder of T3) just published a long letter to OpenAI. He conducted in-depth testing of GPT-6.2 Astra, and his conclusion was quite disheartening:
This model is intelligent, but as a developer, he'd rather use Claude Opus 5.5 every day.
He listed several frustrating problems:
1. Prone to freezing on long tasks. It would suddenly ask the user for internal implementation details that are impossible for them to know, and then just stop. He spent two hours working on an Outer Wilds level, only to get stuck because of a video file naming issue.
2. Extremely poor interpretability. After the model finishes its work, the output summary is full of trivial technical details that even the developers themselves can't understand. He now has to let Astra do the work, then switch to Opus 5.5, and have the latter explain what Astra actually did.
3. Prone to getting stuck in a vicious cycle. He tested a TypeScript project rewritten in Rust. Astra ran it for seven days, then another four, with almost no progress, even causing a 56% performance drop in compiler performance. Opus 5.5, on the other hand, completed the project in less than two weeks, with token costs of approximately $20,000.
4. Lack of a sense of purpose. Two different people, using different games and different hints, asked it to design game levels, and both ended up with themes of water, music, and bells, and both even named it Bellwether.
Share To
Timeline
HotFlash
APP
X
Telegram
CopyLink