OpenAI Astra Explained

Astra is OpenAI's current frontier model, released on September 3, 2026 as GPT-6 Astra. OpenAI first named it in an August 1, 2026 research post on mathematics, then paused parts of the work six days later after a safety evaluation.

It now has the numbers it was missing all summer. GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens, holds a 1,050,000-token context window, writes up to 128,000 tokens in one call, and knows nothing after April 30, 2026.

For the launch timeline, see the OpenAI Astra release date page. The ChatGPT 6 features page covers the naming, and the OpenAI Astra benchmarks page covers the evidence.

What Is OpenAI Astra?

Astra is the frontier model family from OpenAI, announced on August 1, 2026 in a research post about mathematics and theoretical computer science, and released on September 3, 2026 as GPT-6 Astra.

It succeeds the GPT-5.6 lines already in production: Sol, Terra, and Luna. The GPT-5.7 label that pre-launch reporting raised was never used, and the shipped API id is gpt-6-astra.

OpenAI published the specifications at launch that were missing at announcement. There is now a rate card, a context window, and a system card, which is what a buyer needs to size a budget against.

• Announced — August 1, 2026, in an OpenAI research post on mathematics and theoretical computer science.
• Released — September 3, 2026 to a limited set of organisations, with the general release on September 4, so a migration can now be planned against real dates.
• Naming — GPT-6 Astra. The API id is gpt-6-astra, which is the string an integration has to match.
• Positioning — the successor to GPT-5.6 Sol, Terra, and Luna, aimed at long-running work split across several agents.
• Access — ChatGPT Plus, Pro, Business and Enterprise over the days after launch, plus the OpenAI API and AWS.

What Astra Has Already Proven

Astra's one public result is a set of ten solved problems in mathematics and theoretical computer science, published by OpenAI on August 1, 2026. An internal version of the model produced them.

OpenAI put the compute cost for all ten solutions at roughly $2,000 at GPT-5.6 Sol application programming interface (API) rates. That price is why the result travelled well outside mathematics.

The proofs can be checked by machine rather than taken on trust. OpenAI released a 249-page manuscript and Lean 4 proof certificates on GitHub under an Apache 2.0 licence. In Lean, the keyword sorry marks a step nobody has proved. The repository reports a sorry count of zero, so no step in the ten formalised proofs is left unproved.

• An explicit construction of a non-sofic group, a question open since Mikhail Gromov set out soficity in 1999.
• A disproof of Connes's rigidity conjecture on von Neumann algebras.
• A proof of Ehrhart's volume conjecture.
• Three problems from the Erdos catalogue, including problem 183 on multicolour Ramsey numbers.
• Roughly $2,000 of compute for all ten solutions at GPT-5.6 Sol API rates, which made the result a business story rather than only a mathematics one.

Why OpenAI Paused Work on Astra

OpenAI paused parts of its frontier training in August 2026 after an internal evaluation placed Astra at the Critical cybersecurity capability threshold. That is the highest severity level defined in OpenAI's Preparedness Framework, and Astra is the first model OpenAI has put there.

The evaluation was dated August 7, 2026. It followed a July test, reported by Axios, in which an internal model executed 17,600 autonomous intrusion actions against Hugging Face infrastructure.

The response was concrete rather than a statement of intent. OpenAI told Axios on August 18, 2026 that it had stopped two weeks of deployment-focused reinforcement-learning (RL) training and put its largest planned frontier RL run on hold. It also began rewriting the Preparedness Framework, most of which was written in 2023.

• July 2026 — an internal test produced a model that ran 17,600 autonomous intrusion actions against Hugging Face infrastructure.
• August 7, 2026 — a safety evaluation places Astra at the Critical cybersecurity threshold, the highest level the Preparedness Framework defines.
• August 18, 2026 — OpenAI confirms a two-week stop on deployment-focused RL training and a hold on its largest planned frontier RL run (Axios).
• Framework rewrite — the Preparedness Framework itself is being rewritten, because the 2023 text did not anticipate a model reaching this threshold.
• New controls — earlier alignment and security checks, more monitoring during development, and higher safeguards when scaling post-training.
• September 2, 2026 — OpenAI says it is preparing to release Astra with stronger safeguards, and that advanced capabilities go to selected early testers first.

What OpenAI Published About Astra at Launch

OpenAI published prices, a context window, an output ceiling, a knowledge cutoff, and a system card when it released GPT-6 Astra. Those are the figures a buyer needs to size a migration, and none of them existed a week earlier.

The output price is the one that moves a budget. At $50 per million output tokens against $10 per million input, reading a large document set is cheap and writing a long one is not, so the first cost control on an Astra workflow is usually a cap on output length.

The context window changes the architecture rather than the bill. At 1,050,000 tokens a full contract set or codebase fits in one request, which removes the chunking and retrieval layer teams build around smaller windows, and that layer is often more expensive to maintain than the tokens it saves.

• $10 per million input tokens and $50 per million output tokens, so long-form generation is the expensive half of any Astra workload.
• A 1,050,000-token context window, so a document set that used to be chunked can go in as one request.
• A 128,000-token maximum output, which caps a single call and pushes long deliverables into stages.
• A knowledge cutoff of April 30, 2026, so anything later has to arrive through web search or your own retrieval.
• A system card, which documents the evaluations OpenAI ran and the limitations it found.

Is Astra the Same Thing as GPT-6?

Yes. OpenAI released the model as GPT-6 Astra on September 3, 2026, so Astra and GPT-6 name the same thing and the two searches now land on one model.

Astra was OpenAI's designation for the generation, and it stayed in the product name rather than being dropped at launch. The GPT-5.7 label that reporting had also raised would have placed the model inside the existing GPT-5 series, and OpenAI did not use it.

The naming history, and what shipped in ChatGPT alongside it, sits on the ChatGPT 6 features page.

What Astra Changes for Teams Running GPT-5.6 Today

Move one workload, not the stack. At Layer3Labs, we build and run AI systems inside other people's businesses, and a launch week is the worst time to re-platform a pipeline that already works, because the failure modes of a new model show up in your data rather than in its benchmarks.

Long-horizon work is where the price is worth paying. Astra is aimed at problems that run across several agents and many steps, and at $50 per million output tokens it has to be earning something a cheaper tier cannot deliver.

Who should ignore this entirely: any team getting value from GPT-5.6 Terra or Luna on routine work. Ticket triage is the opposite of a long-horizon problem, and a cheaper tier will keep beating Astra on cost there no matter what it scores.

What would change this answer: your own evaluation set showing Astra completing a job the GPT-5.6 tiers fail, or a workload whose documents only fit in a 1,050,000-token window. Either one is a reason to move that workload and nothing else.

• Do now — run your own evaluation set against gpt-6-astra, because launch benchmarks describe mathematics and exploit challenges rather than your documents.
• Do now — add a frontier-capability question to your vendor review, since Astra is the first model OpenAI has classified Critical for cybersecurity capability.
• Do now — cap output length on anything you point at Astra, because output tokens cost five times what input tokens cost.
• Check — whether any workflow depends on reading the model's reasoning, since recurrent depth hides part of it.
• Do not do — re-platform a working GPT-5.6 workflow in launch week on the strength of a benchmark table.

How to use GPT-6 Astra

You do not host GPT-6 Astra yourself — you use it through a tool, so "getting started" really means choosing the right one.

The fastest way to put GPT-6 Astra to work day to day is inside an AI IDE, and Cursor is the most popular — it supports it directly, so you can be working in minutes. The maker's own option is Codex for GPT-6 Astra, if you want the native experience. Prefer a different editor? Windsurf, Zed, and GitHub Copilot drive these models too.