GPT-6 Astra

OpenAI announced GPT-6 Astra on 3 September 2026, first to a limited set of organizations and then across ChatGPT Plus, Pro, Business, and Enterprise, the API, and AWS. Pro, Business, and Enterprise plans also get an Astra Pro variant. The API model ID is gpt-6-astra.

What it costs

$10 per million input tokens and $50 per million output, which is five times what GPT-6 Sol costs. Batch and flex processing run it at half price. Fast mode doubles it. Prompts over 272K tokens are billed at the higher rate for the entire request, so if you can split the work without re-sending the same context twice, two calls under the line can come out cheaper than one over it.

On price alone, Astra is the expensive option and GPT-6 Sol and Luna cover volume work at a fraction of it. The catch I ran into is that the work which genuinely needs Astra rather than Sol is exactly the work that makes it expensive: hard, open-ended prompts where it thinks for a long time. It can burn a serious share of a token budget on one of them.

How I actually use it

Not as a daily driver. Fable is still my favorite, and Astra didn’t close a big enough gap to move me. What it’s good for is a cold second opinion on something important, which is a specific and underrated job — I’ve written about why a second model reads your work better than the one that wrote it.

The other thing worth flagging is tool use. In the desktop app, Astra handles tools better than Claude does at the moment. If your work involves the assistant reaching out to other things rather than just answering, that difference shows up quickly.

The safety story

OpenAI’s system card rates Astra as Critical for cyber capability under its Preparedness Framework, the first of its models to carry that rating. The consequence is that its cyber-sensitive capabilities ship behind a trusted-access program rather than to everyone, and OpenAI put a comparison of how well each model resists jailbreaks, Astra against GPT-5.6 Sol, in the same document.

Not the other Astra

Google DeepMind runs a research prototype called Project Astra, whose capabilities have been flowing into Gemini Live. Different company, different thing, same word. If you see “Astra” in a Google context, it isn’t this model.

The Verdict

Best for: Work where the answer has to be right the first time and the token bill is not the constraint.