What GPT-6 Astra Is, in Plain Terms

GPT-6 Astra is OpenAI's flagship model in this generation: shipped on 2026-09-03, priced at $10 per million input tokens and $50 per million output, advertised with a 1.05M-token context window and a 128K output ceiling, and described by its own vendor as the model for "complex reasoning and coding". It is the earliest of the four models in this comparison to ship, the most expensive per token, and the best-measured of them — 45 of 68 benchmark rows across 7 of 8 categories on the independent coverage board, read 2026-10-07. The GPT-6 Astra API is carried here at list rate with no markup, so everything below can be checked against an invoice rather than a slide.

Astra is the most expensive model here per token and the *second cheapest* per finished task. That inversion is the first thing worth understanding about it.

What actually shipped

Astra is one rung of a three-rung line, not a standalone product. The vendor's own routing advice sends complex reasoning and coding to it, cost-sensitive high-volume work to GPT-6 Luna below it, and everything in between to GPT-6.1 Sol in the middle. Read as a family, the line is a price ladder with a stated default, and Astra is the rung the vendor points hard problems at.

It is also a new name rather than a version bump. The independent record's deprecation chains are explicit about replacements — `gpt-6-sol` resolves to `gpt-6.1-sol`, `claude-opus-5` to `claude-opus-5-5`, `grok-4-6` to `grok-4-7` — and Astra carries no such entry, which is what a fresh name in a family looks like rather than a successor to one specific model.

One caveat on the date, and it is ours rather than OpenAI's: our own catalog lists Astra's release as 2026-09-04, one day after the vendor's changelog. When two records disagree by a day, take the vendor's.

The specification, including the parts you cannot change

Most of Astra's specification sheet reads as a superset of the model below it, which is the point of a flagship.

Three lines in that table are worth a second look.

The context row is the first. The vendor advertises 1.05M and the independent board records a round 1,000,000 — the same pair of numbers Sol has, and the same pattern the board shows for several models, where it reports a vendor's advertised figure as-is in some cases and a round bound in others. Two sources, one of them slightly conservative, and no practical difference: both are large enough that the limit is not what will stop most workloads.

The second is the pricing tier at the bottom, and it is ours rather than the vendor's. On our own rate card, Astra requests above 272,000 input tokens are billed at $20 in and $75 out rather than $10 and $50. That step is OrcaRouter's, not a line from OpenAI's published card, and it only matters to workloads that habitually send very long inputs.

The third is the ladder, because it comes with a restriction. Astra's reasoning settings run from low to max with no `none` rung, and the model is served with fixed sampling — temperature, top_p and logprobs are not exposed as adjustable parameters. If your pipeline depends on sampling a prompt several times at different temperatures and voting, that dependency has to move to a different model or a different design.

What the independent record says

Astra is the best-evidenced of the four models here, which is a claim about the size of the record rather than about the scores in it.

On the independent coverage board read 2026-10-07, it publishes 45 of 68 benchmark rows across 7 of 8 categories, with a composite score of 84.9 and a board rank of 2. The one model above it holds 53 of 73 rows and 8 of 8 categories — enough headroom that the ordering between them is not close.

On the independent evaluation harness, also read 2026-10-07 and measured with the model running at its Max reasoning setting:

Two of those rows deserve calling out. The GPQA figure of 96.06% is the only one on the board for any of the four subjects — the other three carry no row there at all. And the Intelligence Index of 52.67 sits one point above the mid-tier model in the same family and nearly six below the highest score of the ten models on that harness, so "flagship" here means the top of one vendor's line rather than the top of the board.