GPT-6 Astra has moved from limited preview to general availability on the OpenAI API. It is listed at $10 per million input tokens and $50 per million output tokens for short context, unchanged from the announcement price.

What changed

On September 3 Astra was described as a limited preview for trusted partners. It is now on the public pricing page as an available model — a turnaround of days rather than the weeks or months such previews usually take.

One restriction is documented: fast mode is unavailable for GPT-6 Astra under EU data residency. Teams that route requests through EU endpoints for compliance reasons get the model but not the low-latency path.

Why the EU caveat matters

Data residency is not an optional extra for a lot of European organisations — it is the condition under which using a US model provider is permissible at all. A capability that exists everywhere except behind the residency boundary creates a two-tier experience where the teams with the strictest requirements get the slowest service.

It is worth reading as an infrastructure signal rather than a policy one. Fast mode usually depends on specific hardware in specific regions, and matching that footprint inside the EU takes longer than flipping a flag.

Where it sits now

OpenAI's lineup runs Astra at $10/$50 on top, with GPT-5.6 Sol at a promotional $4/$20 through November 21, Terra at $2/$12 and Luna at $0.20/$1.20 beneath. A 50x spread between the cheapest and most expensive model in one family is a wider range than the industry had a year ago.

Regional processing carries a 10% uplift for models released on or after March 5, 2026, so the EU price for Astra is higher than list as well as slower.

What to do

If you are evaluating Astra from Europe, benchmark on the residency endpoint rather than the default one — the numbers you get from a US endpoint will not be the numbers you run in production. For everyone else, the decision is the usual frontier-tier question: is the answer worth $50 per million output tokens on this specific task, or does a model ten times cheaper close the gap.