Quick Answer
GPT-6 Astra (codenamed Astra / mewfour, pretraining project "Doug") is OpenAI's next flagship model — and it is not released yet. It's rumored to be a ~10-trillion-parameter Mixture-of-Experts model and the first frontier model to train multi-agent coordination natively from pretraining. It surfaced on August 1, 2026 via a 249-page paper showing it solved 10 hard math problems, then hit an internal delay after its cybersecurity eval reached "Critical."
How Astra Was Revealed
OpenAI didn't launch Astra with a press release. On August 1, 2026, it published a 249-page paper demonstrating the model solving 10 math problems previously considered human-level difficulty, at roughly $2,000 in tokens per problem, with each proof formally verified by the Lean 4 proof assistant. A week later, on August 7, its cybersecurity capabilities were rated Critical, and OpenAI slowed the release into isolated testing.
Known Facts vs. Rumors
| Item | Status | Details |
|---|---|---|
| Codename | Confirmed | Astra / mewfour (project Doug) |
| Size | Rumored | ~10T parameters, MoE |
| Multi-agent | Confirmed | Natively trained from pretraining |
| Math ability | Confirmed | 10 problems, ~$2,000 each, Lean 4 verified |
| Release date | Unconfirmed | Prediction markets: ~59% by Sep 15, ~72% by Sep 30 |
| Pricing | Unconfirmed | Likely above GPT-5.6 Sol ($5/$30) |
| Final name | Unconfirmed | Possibly "GPT-6" or a GPT-5 variant |
The key caveat: the "10T parameters" figure is rumor, not an official number. The only hard facts are native multi-agent training and the 10 verified math proofs.
What It Means for Developers
- Reasoning paradigm — multi-agent natively means one prompt can internally spawn multiple cooperating agents; task decomposition and aggregation happen inside the model.
- Cost — $2,000 per math problem signals extremely dense inference; running Astra on everything is economically unrealistic.
- Access — it will speak the OpenAI-compatible API, so migration from GPT-5.6 is a
base_url+ model-name change.
FAQ
Q: Are GPT-6 and Astra the same thing? Probably, but OpenAI hasn't finalized the name. Astra/mewfour are internal codenames; the launch name could be "GPT-6" or a new GPT-5 tier.
Q: Can I use it today? No. No release, no waitlist, no date. Any "GPT-6 API available now" claim is third-party hype — verify before trusting.
Q: What is it actually best at? The only hard evidence is 10 formally-verified math proofs and native multi-agent. There are no official coding or general benchmarks yet — don't trust "beats everything" claims.
Summary
GPT-6 Astra is the most-discussed unreleased model in AI right now: a natively multi-agent, reasoning-heavy flagship with formally-verified math ability, delayed by a cybersecurity review. To use it on day one, prep your OpenAI-compatible access layer now — TeamoRouter is one key that will cover GPT-6 Astra, Claude, and DeepSeek.
Get Started
Sign up for TeamoRouter and set base_url to https://api.teamorouter.com/v1 — swap the model name to gpt-6 when Astra ships.