← Back to articles

NewsBy: openaileaks

OpenAI is quietly feeding Codex users an unreleased GPT-6.2

Some Codex users found a model called gpt-6.2-astra in their usage stats, though officially it doesn't exist. OpenAI appears to be testing it on live traffic.

OpenAI is quietly feeding Codex users an unreleased GPT-6.2
Photo: TestingCatalog

Over the weekend, users of Codex, OpenAI's coding agent, spotted an odd line on the page that tracks their plan usage. Next to the familiar GPT-6 Astra, released in September, sat a model called gpt-6.2-astra. Officially there is no such model: OpenAI hasn't said a word about it, its docs still list only GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, and the model picker still reads GPT-6 Astra.

Going by screenshots that spread on X on November 1, the new model shows up only on the October 31 bar. The share varies a lot: 1.6% of one user's week, 19.8% for another, 37.7% of a single day for a third. Someone else posted a raw response event with the model field reading "gpt-6.2-astra". It looks like some requests are being quietly routed to the new model, either as an A/B test or the start of a staged rollout. People simply don't know they are talking to a different model.

Testers who believe they got 6.2 are already showing off. One shared the classic pelican on a bicycle as a 3D scene you can spin 360 degrees. Another says it built a semi-open-world 3D RPG from a single prompt and asks whether it can catch Claude Opus 5.5. There is no way to confirm which model actually answered, because the interface doesn't say.

The version number is its own mystery. GPT-6.1 Astra never shipped: reportedly OpenAI shelved it in late September after internal testing raised concerns about deception and scope. Jumping straight to 6.2 could mean a fixed, retrained model, but that is speculation, and OpenAI hasn't confirmed anything. Meanwhile the company released GPT-6.1 Terra just a couple of days ago, and Google and Anthropic are lining up their own moves, so November looks busy.

For now this is not an announcement, it's a leak through usage stats. But if the model is already handling live Codex traffic, a public launch may be only days away.

original