OpenAI launches GPT-6 Astra, its most capable model

What's new

On September 3, 2026, OpenAI launched GPT-6 Astra, which the company calls its most capable and best "aligned" (that is, most controlled and reliable) model to date. The model is rolling out first to a select group of organizations and will then come to all users of ChatGPT Plus, Pro, Business and Enterprise, as well as via OpenAI's API, Microsoft Azure and Amazon AWS. Astra has a memory window (context length) of up to 1,050,000 tokens, or roughly a million pieces of text, and can respond with up to 128,000 tokens in one go.

What's clever

Astra is the first OpenAI model to reach what the company calls a "critical" level in cybersecurity. This means the model can on its own find unknown security holes in well-protected systems and build working attack methods without step-by-step help from humans, something that previously required experienced experts. Precisely for that reason, OpenAI has delayed parts of the rollout to tighten safety guardrails, and the most advanced cyber capabilities are only available to a select group of organizations in the Daybreak security partnership. For everyday use, Astra is meant to be especially good at long, independent tasks where it uses a computer and browser on its own, for example research, coding and data analysis.

Cheaper or better?

In terms of price, Astra is more expensive than its predecessor. Via the API, it costs $10 per million input tokens and $50 per million output tokens, about 2.5 times as much as GPT-5.6 Sol. For regular ChatGPT users, there is no extra charge; it is included in the Plus, Pro, Business and Enterprise subscriptions, although with a cap on the number of messages on some plans. On benchmarks, the picture is mixed. On advanced mathematics (FrontierMath), Astra scores 97.6 percent versus 87.8 percent for the Claude model Fable 5.1, a clear leap. On a test of independent terminal work (Terminal-Bench), Astra comes in at 57.7 percent versus Claude at 55.8 and Google's Gemini 3.8 Flash at just 19.1 percent. But on an overall intelligence measure (Artificial Analysis Intelligence Index), Claude's model actually scores higher than Astra, so OpenAI is not ahead on every point.

What it's good at

Astra is built for demanding professional work: advanced analysis, software development, deep research, scientific work and documents, as well as long tasks where the AI itself controls the computer and browser over time. On the other hand, the model deliberately refuses to help with the most dangerous cybersecurity tasks, such as building ready-made exploit code against known vulnerabilities.