Astra (OpenAI)

OpenAI's "next major model". ~~Unreleased.~~ Shipped, or at least reachable: as of 2026-09-04 it is on the API with published rates — see What changed on 2026-09-04 below. Named publicly on 2026-08-01 in the Ten advances in mathematics and theoretical computer science publication, which credits an internal version of Astra with ten results in mathematics and TCS. That announcement is currently the only first-party statement this wiki holds about the model, and it is not a model announcement — it is a mathematics paper that happens to name its author.

Canonical node for Astra. The result itself, its Lean formalizations, and what the certificates do and don't establish live in ../research-wiki as ten-proofs and lean-certificate — routed there on 2026-08-03 because the subject is theorem proving; this page is the model-market half.

What OpenAI actually stated

"The results were achieved by an internal version of Astra, our next major model. The total number of tokens needed to find solutions to these problems would cost roughly $2,000 at Sol API rates. These arguments were then prepared into manuscripts by humans with the same model. Afterward, the model formalized each argument in a Lean certificate."

Four claims, all first-party: it found the arguments, a human-plus-model pass wrote them up, it wrote the Lean, and the search cost about $2,000 of tokens priced in another model's currency.

The $2,000 figure is a compute proxy, not a price

Read carefully, the cost sentence prices Astra's run at Sol API rates — Sol being the GPT-5.6 flagship that actually ships. ~~Astra's own pricing does not exist publicly, so~~ (corrected 2026-09-05: it does, as of 2026-09-04 — $10/$50; see the section below. The rest of this paragraph stands, because the quoted sentence was written when it did not.) The number is a token-volume proxy dressed as a dollar figure: it says roughly how much generation the ten results took, expressed in the nearest published rate card. Anyone reading it as "ten open problems for two thousand dollars" is reading a price for a product that, at the time of writing, had no price — and at Astra's own card the same volume is $3,300–$4,000.

Useful anyway, and unusually so. Most capability announcements in this wiki give no cost at all (claude-opus-5's ratio-only benchmarks are the recent case). This one gives an order of magnitude for the inference budget behind a frontier result, which is the input the cost-per-task argument has been missing.

What changed on 2026-09-04

astra-pelican-grid (simon-willison, T2) is the first non-first-party contact this wiki has with the model, and it settles two of the three questions this page was carrying.

It has a price. $10 per million input, $50 per million output, against Sol's $5/$30 (the GPT-5.6 Sol rate as Willison gave it on 2026-09-04; OpenAI's page read on 2026-09-23 lists GPT-5.6 Sol at a promotional $4/$20, and the new GPT-6 Sol at $2/$10 — openai-api-pricing-docs, gpt-6) — 2× on input, 1.67× on output. Willison adds the qualification that matters more than the ratio: Astra "uses significantly less tokens at each of the levels, making the prices at the different levels closer than they might otherwise be." So the rate card overstates the cost gap, and his worked case is Astra at low reasoning beating every Sol level at any reasoning, for 9.55 cents.

Which re-prices the $2,000 figure below. That number was OpenAI's token volume expressed at Sol rates. At Astra's own card the same volume costs between 1.67× and 2× as much depending on the input/output split — call it $3,300–$4,000 (our arithmetic, and the split is unpublished, so the bounds are the honest form). The $2,000 was never Astra's price and now demonstrably is not.

It is a reasoning model with no off switch. Reasoning levels run low / medium / high / xhigh / max, and reasoning=none is unsupported, where the GPT-5.6 line supports it.

Quality, judged by eye and worth what that is worth. Every Astra pelican from low to xhigh beats the best Sol pelican; Astra at max is "really good"; and below max it "doesn't reliably get the pelican legs on both sides of the frame." One prompt, one sample per cell, no rubric — not a measurement, and the first outside comparison of any kind this page has had.

And a hypothesis about lineage, held as a hypothesis. On the same prompt, Astra and Luna both consumed 16 input tokens while Sol and Terra consumed 26; Willison asks whether "Astra and Luna are more related to each other than OpenAI let on." That speaks to the third open question below. It is one prompt and two numbers — consistent with a shared tokenizer, which is consistent with several stories — so it stays a question until something else supports it.

What is not known

~~Everything else. No parameter count, architecture, context window, modality list, release date, availability, pricing, or safety documentation.~~ Updated 2026-09-05: availability and pricing are now known (above). Still unknown: parameter count, architecture, context window, modality list, and safety documentation. No benchmark table — no SWE-bench, GPQA, HLE, ARC, nothing this wiki's llm-benchmarks page is built to compare. No system card. It is not clear whether the "internal version" that did the mathematics is the version that will ship.

The Sol/Terra/Luna tiering of GPT-5.6 (openai-gpt56-grok45-clash) says nothing about whether Astra follows the same split, and OpenAI has not said whether Astra succeeds GPT-5.6 or sits beside it.

Why this page is unusual for this wiki

Every other model page here rests on numbers the vendor produced and nobody else can re-run. Astra's public evidence is the opposite shape: ten machine-checkable proofs in a public repo, verifiable by a stranger with Lean and a laptop, with no access to the model required. See lean-certificate.

That inverts the usual reliability problem without solving it. What the certificates establish is that the mathematics is correct. They establish nothing about the model — not that it found the proofs the way the published narrations describe, not how many attempts it took, not what human steering went in, and not that a released Astra will do anything comparable. The verifiable artifact and the capability claim are two different objects, and only the first one is checkable.

So: strong evidence that ten theorems are true, weak evidence about a model nobody outside OpenAI can run.

Open

Does a shipped Astra resemble the internal one? Still open, and now sharper rather than closer: a public endpoint exists, so the comparison is finally possible, and nobody has run the ten mathematics problems against the shipped model. That is the experiment this question has been waiting for and it is now buyable.
~~Pricing, and whether $2,000-class inference budgets are purchasable.~~ Closed 2026-09-05. $10/$50 per million, so a $2,000-at-Sol-rates run costs roughly $3,300–$4,000 at Astra's card. The budget is purchasable by anyone with the money; whether it buys the same result is the question above.
Where it sits against GPT-5.6. Still open, with one weak signal added: astra-pelican-grid's identical 16-token input counts for Astra and Luna against 26 for Sol and Terra. Successor, sibling, or a different line entirely — a tokenizer coincidence on one prompt does not decide it.
Nobody has measured it. New 2026-09-05. The only outside contact is one person's SVG grid, judged by eye. No independent benchmark, no system card, no third-party evaluation of a model that is now on sale. Retracted 2026-09-16 (claim drain). The model launched as GPT-6 Astra on 3 September 2026 (1M-token context, $10/$50 per million, so the pricing above held), and the independent numbers arrived the same week. Artificial Analysis puts it at 61.2 on its Intelligence Index v4.1.1, behind Fable 5.1 (65.7), Opus 5 (63.1) and Fable 5 (62.1), and second on its Coding Agent Index (67.0 to Fable 5's 68.1). The ARC Prize Foundation's verified run is the instructive one: 62.7% on the standard provider-neutral harness, 99.9% on OpenAI's own adapter harness, and the launch chart showed the higher figure without noting that older models were not run under the same adapter. Irregular reports 86 of 226 FrontierCyber challenges solved against 34 for Sol. Fortune (4 Sep) documented OpenAI editing its own launch metrics after publication — the hallucination rate went 4.2% → 2% → 4.2% across six snapshots in one afternoon, and competitors' FrontierMath Tier 4 scores moved twice — and says a system card exists but explains little of the internal hallucination benchmark; Vellum's explainer (same week) still calls the card forthcoming. Both readings are recorded; the card's status is the residue. Sources: Vellum (vellum.ai/blog/gpt-6-astra-benchmarks-explained), DataCamp (datacamp.com/blog/gpt-6-astra), Fortune 2026-09-04. The vendor's own table, for the record: OSWorld 2.0 72.6%, Terminal-Bench 4.0 57.7%, FrontierMath Tier 4 97.6%, ExploitBench 100%. Read with the ARC harness split before trusting any of them.
The ten-problem experiment is still unrun (carried from the item above, 2026-09-16): every independent number is on a general benchmark; nobody has pointed the shipped model at the ten theorems.

astra-pelican-grid · simon-willison · openai · openai-gpt56-grok45-clash · llm-benchmarks · llm-api-pricing · claude-opus-5 · ten-proofs · lean-certificate