Last updated: 2026-09-07 — GPT-6 Astra is now released (September 3, 2026). This article has been updated from release-rumor coverage to shipped-fact coverage; the original pre-release history is kept below for context.
Quick Answer
GPT-6 Astra (codenamed Astra / mewfour, pretraining project "Doug") is OpenAI's new frontier flagship, released on September 3, 2026. The API model ID is gpt-6-astra. Headline specs: a 1.05M-token context window, $10 per million input tokens / $50 per million output tokens, native multi-agent coordination, and state-of-the-art computer use. It succeeds GPT-5.6 Sol — and it is not "GPT-5.6 Astra"; Astra is the GPT-6 generation.
Confirmed Specs (Shipped)
| Item | Status | Details |
|---|---|---|
| Release date | Confirmed | September 3, 2026 |
| API model ID | Confirmed | gpt-6-astra |
| Context window | Confirmed | ~1,050,000 tokens (long-context input tiers apply beyond ~272K; reports say past input is rebilled at double rate — check the official pricing page for current tiers) |
| Pricing | Confirmed | $10 / $50 per 1M input / output tokens; cache reads $1.00, cache writes $12.50 |
| Architecture | Confirmed | Native multi-agent coordination trained from pretraining; parameter count still unofficial (the "~10T MoE" figure remains a rumor) |
| Computer use | Confirmed | Hosted computer-use model; ~72.6% on OSWorld per third-party coverage |
| Availability | Confirmed | API, plus ChatGPT Plus / Business / Pro / Enterprise |
The key caveat carried over from the pre-release coverage: the "10T parameters" figure is still rumor, not an official number. OpenAI has not published a parameter count.
How Astra Was Revealed (the Backstory)
Astra didn't launch with a press release. On August 1, 2026, OpenAI published a 249-page paper showing the model solving 10 math problems previously considered human-level difficulty, at roughly $2,000 in tokens per problem, each proof formally verified by the Lean 4 proof assistant. A week later, on August 7, its cybersecurity capabilities were rated Critical, and the release slipped into isolated testing. It cleared that process and shipped on September 3 as what OpenAI calls its "most intelligent and aligned model yet."
Using GPT-6 Astra
Astra speaks the standard OpenAI-compatible API, so existing GPT-5.6 code carries over with a model-name change:
from openai import OpenAI
client = OpenAI(api_key="YOUR_KEY", base_url="https://api.openai.com/v1")
resp = client.chat.completions.create(
model="gpt-6-astra",
messages=[{"role": "user", "content": "..."}],
)
Practical notes for developers:
- Model ID: use
gpt-6-astra. Older snippets that saymodel="gpt-6"predate the launch naming. - Long context: the window is ~1.05M tokens, but pricing is tiered — short-context and long-context requests bill differently, and requests beyond roughly 272K input tokens are reported to rebill at double rate. Keep prompt bloat out of the expensive tier when you can.
- Caching: cache reads are $1.00/M (10x cheaper than fresh input), cache writes $12.50/M. Prompt caching pays off quickly on Astra's input price — structure repeated system prompts to hit the cache.
- Codex integration: with Astra, Codex gains the ability to preserve and retrieve context when the window fills, which changes how long agent sessions behave in practice.
- Computer use: Astra is OpenAI's strongest computer-use model — screenshot → action loops for desktop automation. Budget accordingly; these loops are token-hungry.
GPT-6 Astra vs GPT-5.6 Sol: Don't Confuse the Names
A common search — "gpt 5.6 astra" — mixes up two different generations:
| GPT-5.6 Sol | GPT-6 Astra | |
|---|---|---|
| Generation | GPT-5.6 (previous flagship) | GPT-6 (current flagship) |
| Input / output per 1M | $5 / $30 | $10 / $50 |
| Context | Smaller | ~1.05M tokens |
| Multi-agent | No (single-model reasoning) | Yes (native, from pretraining) |
| Computer use | Basic | State of the art (~72.6% OSWorld) |
"Astra" is not a GPT-5.6 variant. Sol remains available as the cheaper predecessor; Astra is the new top of the stack at roughly double Sol's list price. For the full cost breakdown, see GPT-6 Astra vs GPT-5.6.
Where to Go Deeper
- GPT-6 Astra benchmarks — every published score, and which independent numbers disagree with OpenAI's table
- GPT-6 Astra API access — keys, base URL, and gateway setup
- GPT-6 Astra pricing — full price math and cost routing
- Codex/Cursor setup for GPT-6 Astra
- GPT-6 Astra vs Claude Fable 5.1 — the flagship-vs-flagship decision
FAQ
Q: Are GPT-6 and Astra the same thing?
Yes. Astra/mewfour were internal codenames; the shipped product is named GPT-6 Astra, with the API model ID gpt-6-astra.
Q: Is "GPT-5.6 Astra" a real model? No. GPT-5.6's flagship tier is Sol. Astra belongs to the GPT-6 generation. If a tool or article lists "GPT-5.6 Astra," it's a naming error.
Q: Can I use it today? Yes — it shipped September 3, 2026, via the OpenAI API and ChatGPT Plus/Business/Pro/Enterprise tiers. It's also reachable through OpenAI-compatible gateways; Moyi API carries it with one key alongside Claude and DeepSeek.
Q: What is it actually best at? Confirmed strengths: formally-verified math (10 Lean 4 proofs at reveal), computer use, and long-context agent work. Independent coding results are mixed — see the benchmarks article before switching your whole workflow over.
Summary
GPT-6 Astra shipped on September 3, 2026 as OpenAI's new flagship: gpt-6-astra on the API, 1.05M-token context, $10/$50 per million tokens, native multi-agent reasoning, and best-in-class computer use. At roughly double GPT-5.6 Sol's price, the smart move is routing it to the tasks that deserve it. Sign up for Moyi API and set base_url to https://api.moyiapi.com/v1 — Astra, Claude, and cheaper fallbacks behind one key.