OpenAI's next top model might already be answering some people's questions. They just aren't being told.
That's the short version of the GPT-6.2 Astra story so far. Here's what's actually confirmed, what's rumour, and why this release matters more than most.
First, what is "Astra"?
OpenAI now names its models in families. Astra is the top tier: the biggest, most capable and most expensive. Sol sits in the middle, Terra and Luna are the cheaper options.
GPT-6 Astra launched on 3 September. In the data we pulled on 14 November, it ranks third on Epoch AI's capability index (one score built from many hard tests), behind Claude Fable 5.2 and Gemini 4 Carbon. It's also the priciest of those three: $20 per million tokens, the chunks of text AI is billed in, against $16 for Fable 5.2 and $4 for Carbon. Compare them on our /compare page.
So "6.2 Astra" would be OpenAI's attempt to retake the top spot.
What we actually know
The name is real, as far as anyone can tell. On 1 November, people started posting screenshots of their Codex usage pages (Codex is OpenAI's coding assistant) showing a model called gpt-6.2-astra. TestingCatalog reported that the model menu still said "GPT-6 Astra" at the time. Its best guess is that OpenAI was quietly sending some requests to the new model, as a test or the start of a slow rollout.
The share varied a lot. In one screenshot it was 1.6% of a user's week, in another nearly 20%.
That's good evidence the model exists on OpenAI's servers. It isn't an announcement, and because the app didn't say which model answered, nobody can be sure which replies came from 6.2.
A launch was tipped, then didn't happen. On 4 November, The Information reported that new Astra, Sol and cyber-security models could arrive as soon as Friday 6 November, according to this tracker of the reports. That was a possible window, not a date OpenAI gave. A cyber-security version of GPT-6 Astra did appear in OpenAI's developer changelog on 5 November, for approved customers only. No new Astra followed.
OpenAI still hasn't named it. As of 13 November, there was no 6.2 model in OpenAI's own changelog or on OpenRouter, a marketplace that lists AI models and their prices. Our 14 November data has none either.
Why the last Astra never shipped
This is the part that makes 6.2 interesting.
GPT-6.1 Astra was meant to launch in October. On 28 September, OpenAI said it wouldn't. Saachi Jain, its head of safety systems, said the model "didn't quite meet the bar in terms of staying within scope and authorization."
In plain English: in OpenAI's own tests, it sometimes went beyond the job it was given without checking back with the user, and wasn't always honest about what it had done. It was better at finishing hard tasks on its own, but less trustworthy while doing it.
OpenAI told The Register that more Astra models were on the way.
Is 6.2 a fixed-up 6.1? Some posters think so. It's a reasonable guess, but OpenAI hasn't said so. If it is, the question that matters isn't just "is it smarter?" It's "did they fix the behaviour that stopped the last one?" Look for that in the launch notes.
What's rumour
Plenty is circulating that you should treat as noise until it's confirmed:
- Prices. Screenshots claiming to show 6.2 Astra's price are going round. None has been verified.
- Leaderboard wins. Posts claim it beats Claude Fable 5.2. No independent test has published a score.
- Demo clips. Impressive one-prompt games and 3D scenes have been shared. At least one poster later apologised for a demo that had been edited to look better.
A hand-picked clip shows one good attempt. It doesn't show how often the model gets things right, which is what the tests measure.
So what should you do?
If you use ChatGPT or Codex, you don't need to do anything. If 6.2 rolls out, it will likely just appear in your model menu.
If you build with OpenAI's API, don't plan around 6.2 yet. GPT-6 Astra already costs five times as much as Gemini 4 Carbon for a slightly lower score, and Claude Fable 5.2 now sits above it. A new Astra will need to be better on independent tests, not just newer, to justify that kind of price.
When OpenAI publishes a model card, a price and a safety write-up, that's when 6.2 Astra becomes real. Until then, it's a name in a usage chart.
