OpenAI has announced GPT-6 Astra, the first model in a new generation and now the top of its lineup. It appears on OpenAI's pricing page at $10 per million input tokens and $50 per million output tokens, with a 1M-token context window and the API model ID gpt-6-astra.
What was announced
The model started as a limited preview for trusted partners rather than a general release. OpenAI positions it as a generational leap in cybersecurity, software engineering, professional work and science — a framing that puts capability and safety in the same sentence, which is unusual for a launch announcement.
The release also followed a delay. After a security incident in July 2026, OpenAI held the model back to add further safeguards before shipping.
What the price says
At $10/$50 Astra sits at the same list price as Claude Fable 5 and ten times above OpenAI's own GPT-5.6 Terra. That gap is the real signal: the frontier tier and the workhorse tier have separated into different products with different economics, rather than one ladder where each rung is modestly more expensive than the last.
The GPT-5.6 family stays in place underneath — Luna at $0.20/$1.20, Terra at $2/$12, and Sol now at a promotional $4/$20 through November 21, 2026. Astra is not a replacement for those; it is a separate, much more expensive class.
Why it matters
For most production workloads nothing changes on day one. A model at $50 per million output tokens is not what you put behind a high-volume classification pipeline. It is what you reach for when a single answer is worth real money — security review, complex refactors, research-grade analysis.
The more interesting question is what it does to the tier below. Every previous frontier launch has pushed the older flagship down in price within a quarter or two. If that pattern repeats, the practical effect of Astra will be felt by teams who never call it directly.
What to watch
Two things worth tracking: when the limited preview opens to general availability, and whether the safety framing translates into published evaluation results. A vendor claiming a leap in cybersecurity capability is also claiming a leap in cybersecurity risk, and the useful version of that claim comes with numbers.