OpenAI Astra: The Preview Field Note
On 1 August 2026, OpenAI previewed its next major model family — Astra — not with a keynote or a launch page, but by pushing ten Lean-4 formalized proofs of previously unsolved problems in mathematics and theoretical computer science to a public GitHub repo. There is no API, no pricing, no ChatGPT surface, and no release date. Every "Astra vs GPT-5.6" or "Astra vs Claude" buying guide you'll read this week is speculation dressed up as reporting.
This page is the honest field note for a Claude user: what's actually confirmed, what's not, and whether it should change anything you're doing today.
- Learn the exact announcement mechanism and what OpenAI did (and did NOT) say on 1 Aug 2026
- Read the ten problems Astra solved and understand why math/TCS was chosen as the reveal surface
- Separate confirmed facts (multi-agent, long-horizon, internal only) from speculation (context size, pricing, ChatGPT rollout, benchmarks)
- Decide whether to keep building on Claude Opus 5 / Fable 5 today, or wait — with a concrete rule of thumb
The one-sentence version
Astra is OpenAI's internal-only, multi-agent, long-horizon next-generation model, revealed via ten Lean-checkable math proofs; it is not the same thing as GPT-5.6 Sol / Terra / Luna, and it has no product surface — so any migration decision you make this month is still a decision about Claude vs GPT-5.6 vs Gemini 3.6, not about Astra.
What OpenAI actually said (and didn't)
The disclosure is unusual: OpenAI announced its "next major model" in the third paragraph of a research post, with no press release, no benchmarks, no availability, and no pricing. That framing tells you as much as the model does — it says "not yet a product, still under regulatory review."
Confirmed on 1 August 2026
- Astra is a model family name. OpenAI describes it as "our next major model." It sits alongside — not inside — the Sol / Terra / Luna GPT-5.6 tiers.
- The results came from an internal version. Not the ChatGPT app, not the API, not a partner-only checkpoint.
- The reveal surface is math + TCS. Ten previously unsolved problems, each with a Lean-4 proof pushed to
github.com/openai/ten-proofsunder Apache-2.0. - The system is multi-agent and long-horizon. OpenAI's own framing is that Astra is "designed to coordinate multiple agents working together on long-running tasks over hours or days," aimed at research, coding, and scientific problems.
- Cost anchor (not model pricing). OpenAI wrote that the tokens used to generate the ten solutions "would have cost about $2,000 at Sol's API rates." That is not Astra's list price, per-proof cost, or reproducibility guarantee — it's a token-volume anchor priced against a different model's meter.
- Government review path. Sam Altman demoed the system to U.S. federal officials in Washington, D.C. in late July 2026. Reporting positions Astra as likely to be the first model submitted under the new U.S. framework requiring federal review before public release.
- Named quote. OpenAI researcher Noam Brown called Astra "a major step for scientific reasoning" and noted, "sadly, no Millennium Prize Problems (yet)."
Not confirmed, despite what you may read
- No release date. No ChatGPT rollout schedule, no API GA, no beta signup.
- No pricing. Not per-token, not per-run, not per-tier.
- No context window size, no max output.
- No multi-modal claims. Nothing about image, audio, or video is asserted.
- No head-to-head benchmarks. No SWE-Bench, no ARC-AGI, no Frontier-Bench numbers.
- No answer to "is it GPT-6?". OpenAI has not said whether Astra ships as GPT-6, as another GPT-5.x point release, or under its own name.
- No confirmation of the tier structure. The Sol/Terra/Luna pattern from GPT-5.6 has not been mapped onto Astra.
If a headline gives you any of the items above as fact this week, the article is speculating.
The ten proofs, in one screen
The choice to reveal Astra through Lean-4 formalized proofs of open problems is doing a lot of work: it lets OpenAI make a capability claim that is machine-checkable, not a leaderboard number that could be gamed. If the proofs verify with lake build All, the mathematical claim is real regardless of what you think of the model.
| # | Problem | Field |
|---|---|---|
| 1 | Improved asymptotic upper bounds on sphere-packing density | High-dimensional geometry |
| 2 | Exponentially stronger upper bounds for binary codes | Coding theory |
| 3 | A construction of a non-sofic group | Group theory / operator algebras |
| 4 | A counterexample to Connes's rigidity conjecture | Operator algebras |
| 5 | New lower bounds for computing the permanent | Arithmetic circuit complexity |
| 6 | Exponential parallel repetition for arbitrary finite games | Quantum complexity |
| 7 | Polynomial-factor hardness of approximation for the Closest Vector Problem | Lattice cryptography |
| 8 | The sharp maximum volume in every dimension (Ehrhart) | Discrete geometry |
| 9 | A superexponential lower bound for multi-color Ramsey numbers | Extremal combinatorics |
| 10 | Counterexamples to the compactness and degeneracy conjectures | Extremal number theory |
Why math + TCS was the reveal surface
Three practical reasons — none of them mysterious once you notice them:
- Machine-checkable claims can't be spun. A benchmark number is a marketing pixel; a Lean proof either type-checks or it doesn't. This is the cleanest possible capability statement in a year where every lab's claims are being questioned.
- No leaderboard-gaming accusations. Nobody suspects Astra of memorizing solutions to problems that were previously unsolved.
- Regulatory optics. OpenAI is walking Astra through the new U.S. federal review process. "Our new model advanced open questions in cryptography and quantum complexity" is a very different framing from "our new model got 96% on SWE-Bench."
What this changes about your Claude vs GPT-5.6 decision today
Nothing that requires action this month. Astra has no product surface: you cannot build on it. Your live tradeoff is still the one from Choosing a model:
- Real-repo software engineering, large refactors, long tool loops on a well-managed platform → Claude Opus 5 (
claude-opus-5) or Fable 5 (see Opus 5 field guide, Fable 5 / Mythos 5 flagship). - Long-context high-volume work at the cheapest viable tier → GPT-5.6 Luna at $1 in / $6 out with the full 1.05M context (Sol / Terra / Luna).
- Browser-heavy or agent-index-shaped tasks → GPT-5.6 Sol or Gemini 3.6 Flash (Gemini 3.6 Flash family).
- Long-horizon multi-agent research or engineering — the exact shape Astra is aimed at — the closest available today is the native multi-agent APIs beta on the OpenAI Responses API, and Anthropic's managed agents + memory stores. Both are shipping products; Astra is not.
Should you wait for Astra?
- Astra has no API, no pricing, and no publicly stated release window. The disclosure is a research post, not a product announcement. Ship on what exists; migrate later if the price/quality is worth it.
- Astra's stated shape — 'multiple agents coordinating on long-running tasks over hours or days' — will be a moving target. Keep prompts, tools, and orchestration behind clean interfaces (see porting prompts across models), so you can drop a new provider in without rewriting your app.
- The ten proofs are a meaningful signal for exactly one class of task: rigorous mathematical/formal reasoning. If that's your workload, subscribe to OpenAI's research page and re-evaluate the day a preview or API opens.
- You cannot buy a SLA on a research disclosure. Frontier previews slip; regulatory review can add months. Revisit when there is a pricing page, a model card, and a rate limit.
Verify the ten proofs locally
# Requires elan (Lean toolchain manager) # Install: curl -sSf https://raw.githubusercontent.com/leanprover/elan/master/elan-init.sh | sh git clone https://github.com/openai/ten-proofs.git cd ten-proofs lake exe cache get # pre-built mathlib artifacts lake build All # type-checks every proof; exits 0 on success
If lake build All exits 0, every one of the ten proofs type-checks against mathlib. That is the strongest capability claim OpenAI has published this year and it is verifiable on your laptop.
Three things journalists keep getting wrong
Where AILmanac will update this page
We will bump lastVerified on this page only when OpenAI publishes a model card, an API endpoint, a pricing sheet, or an official ChatGPT rollout note — not on rumor. Meanwhile, the ten proofs and the multi-agent framing are the durable facts to remember.
- When there is a pricing page → this page becomes a full Astra field guide with a tier table, in the same shape as Sol / Terra / Luna and Opus 5.
- When there is an API → we will add a "port your Claude Opus 5 prompt to Astra" section, alongside our existing porting prompts across models guide.
- If Anthropic ships a comparable "long-horizon multi-agent" primitive → we will update managed agents and cross-link.
Check yourself
0/3Sources & further reading
- Ten advances in mathematics and theoretical computer science — Lean formalizations (
openai/ten-proofson GitHub, Apache-2.0) — the primary artifact: ten Lean-4 proofs you can type-check locally. - OpenAI announces its "next major model" Astra by dropping ten previously unsolved math solutions (The Decoder) — the clearest technical write-up of the 1 Aug 2026 disclosure.
- OpenAI smuggled the announcement of Astra into a blog post about math (Gizmodo) — the naming-ambiguity story (GPT-6 vs Astra vs GPT-5 variant) and the DC-demo reporting.
- OpenAI Astra: Next Major Model Explained (explainx.ai) — includes the "not confirmed" list that most launch coverage skipped.
- Related AILmanac reading: GPT-5.6 Sol / Terra / Luna field note • Opus 5 field guide • Native multi-agent APIs • Porting prompts across models • Choosing a model.
Next
- Current Models & Pricing — the live pricing table AILmanac keeps sourced to official docs (Astra will be added on the day it has an official price).
- This Month in Claude & AI — the rolling digest where new Astra facts land as OpenAI publishes them.
- Long-running agent harnesses — the frontier writeup on the shape of work Astra is aimed at, from the Claude / Managed-Agents side.