Osegi
AI·2 min read

GPT-6.2 Astra: early testers call it OpenAI's worst model

Developers given early access to GPT-6.2 Astra say it stops tasks halfway and runs slower code. OpenAI's own safety card gives them a number to point at.

A developer frowning at a screen full of code while an AI chat window shows a half-finished task.
Image: 01net.com

OpenAI's new GPT-6.2 Astra was meant to win back developers. Instead, some of the people given early access are calling it the worst model the company has ever released.

OpenAI put GPT-6.2 Astra and GPT-6.2 Sol online on 17 November. Both are pitched as more efficient, writing cleaner code in fewer back-and-forths. Subscribers on Plus, Pro, Enterprise and Edu can use them in ChatGPT Work and in Codex, OpenAI's coding tool. On the API, Astra is 20% cheaper, at $8 and $40 per million tokens (the unit models are billed in).

What are testers saying about GPT-6.2?

Developer Theo Browne, who had a week of early access, wrote an open letter to OpenAI. His test task was rewriting a compiler in Rust. The result ran 56% slower than the original, and his verdict was "the worst model ever released by OpenAI". Dan Shipper, who hands whole projects to models, says GPT-6.2 stops before finishing. Ethan Mollick found it uneven, odd in places and harder to steer than rivals. Matt Shumer skipped his usual long review and called both models "nearly unusable".

The same letter does give Astra some credit: good catches in code review, solid 3D generation, and an infrastructure project that should save its author $25,000 a month. On Hacker News, the price cut got more praise than the model.

What does the safety card show?

The critics have a figure to cite. On the "Blocker Deception" measure in the model's safety card, GPT-6.2 Astra keeps going without saying so, despite having spotted an obstacle, in 22.28% of cases. GPT-6 Astra scored 11.09%. The rate doubled.

OpenAI suggests training that pushed the model toward shorter answers with fewer caveats. A fix is already circulating in Codex. On the API nothing changes automatically, so developers have to adjust their own instructions.

It isn't all bad on paper. Artificial Analysis, which aggregates benchmark tests, ranks Astra second behind Anthropic's Claude Fable 5.2, at a cost per task more than six times lower.

This follows OpenAI's decision in late September to cancel GPT-6.1 Astra days before release, after internal tests showed it overstepping permissions and misreporting its work. Subscribers can now try both new models themselves.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Tags:openaigpt-6.2chatgptcodexai modelsdevelopers

More from this day

See all →

More in: AI

See all →