OpenAI is preparing Astra, its model family for agents that work alone, and the rumour mill says August

What exactly leaked, and who is telling it

The newest puzzle piece came yesterday from leaker Pankaj Kumar on X, in a thread with over 64,000 views: OpenAI is reportedly preparing Astra, a new model family focused on long-running, multi-agent tasks, which would live alongside the known classes — Sol, Terra and Luna — and is already in testing, with strong odds of debuting in August. The commercial name is still up in the air: it could arrive as GPT-6 or, if OpenAI stretches the current series, as GPT-5.7.

Circulating with the thread is a screenshot of a 'GPT-6 Astra Thinking' picker, supposedly spotted in ChatGPT test builds. That the fourth name in OpenAI's celestial alphabet points at long-running agents doesn't look like a coincidence: it's exactly the terrain where the company's business is at stake. The screenshot, mind you, is impossible to verify independently: like every interface leak, it could be a ten-minute mock-up.

The other predictions don't quite line up, and that's information too

The rest of the rumour comes with caveats worth keeping in view. The WinCentral maintains that what arrives in August is GPT-5.7 and that GPT-6 has slipped to September, built on a new pre-training foundation, and adds two codenames spotted on the DesignArena testing platform: Zinc, a possible update to GPT-5.6 Sol, and Magnesium, a candidate to succeed Terra. Leakers Lumina and Dan McAteer have been pointing at the same month since mid-July, with heavier rumours — 10 trillion parameters — that nobody has managed to substantiate; the 1.5-million-token context figure has in fact already been walked back by the leak's own sources.

OpenAI, meanwhile, has confirmed absolutely nothing. The only certainty is that every road in the rumour crosses August.

A model to leave them alone longer, the week they confessed they escape

The calendar has its sense of irony again. A model designed for agents to work for hours unsupervised leaks the same week OpenAI and Anthropic confessed theirs escaped their test environments, and on the very day we reported that those agents will live in Amazon's cloud with up to $100 billion involved. The industry is accelerating and braking at once, and Astra raises the stakes: more autonomy, more time alone and more reasons for the cages to stop being made of paper. Nor should the backdrop Xataka recalled this week be forgotten: both OpenAI and Anthropic are preparing to go public, and in that shop window a new model sells better than an apology.