OpenAI’s next model is in a strange and genuinely new position: officially real, officially impressive, and officially unscheduled. Astra has been named by OpenAI itself, credited with ten mathematical results, demonstrated to policymakers in Washington, and then publicly slowed because OpenAI cannot rule out that it crosses the company’s highest cybersecurity risk threshold.
That combination makes most of what ranks for “Astra release date” unreliable within a week of being written. So this page is a tracker, the same format as our Fable 5.1 tracker: every claim dated, every source named, official statements separated from reporting and both separated from rumor. When the story moves, the update lands here, same day.
One disambiguation before anything else: this is OpenAI’s Astra. Google DeepMind’s Project Astra, the universal-assistant research project from May 2024, is an unrelated product that happens to share the name. If you arrived looking for Google’s assistant, this is not that page.
On this page · 6 sectionsOpen
- Astra is OpenAI’s confirmed next major model: OpenAI first named it officially on August 1, 2026, crediting ‘an internal version of Astra’ with ten results on long-standing open problems in mathematics and theoretical computer science.
- There is no release date. OpenAI said on August 7 it cannot rule out Astra meeting the Critical cybersecurity threshold in its Preparedness Framework, and by August 18 a significant number of Astra workloads remained paused under upgraded security requirements.
- The reported shape, from The Information’s July 31 scoop and OpenAI’s own statements since: multiple agents collaborating on long-running tasks, major gains in agentic coding, and training aimed at enabling persistent agents.
- The name may not survive: The Information reported OpenAI had not decided whether Astra ships as GPT-6, a GPT-5-series release, or something else. ‘GPT-6 Astra’ is a search phrase, not a product.
- Astra was NOT the model behind the Hugging Face incident - OpenAI’s own technical report attributes that to a different internal model, though one from the same model family with different post-training.
- Astra already has one confirmed non-customer: OpenAI’s August 28 Cursor announcement states future models, including Astra, will not be supplied to Cursor after its SpaceX acquisition.
- This is a living page: every claim below carries its date and source, and updates land here same day as the story moves.
- What is OpenAI Astra?
- OpenAI’s next major model, officially named on August 1, 2026. Reporting and OpenAI’s statements describe a system built for long-running, multi-agent work, with major advances in agentic coding, cybersecurity, and mathematical research. It is not released and has no announced date.
- When will Astra be released?
- Unknown. OpenAI has never announced a date, and on August 7, 2026 it slowed Astra development after concluding it could not rule out Critical-level cyber capability. As of late August, a significant number of Astra workloads remained paused.
- Is Astra GPT-6?
- Undecided, per The Information’s reporting: OpenAI had not settled whether Astra ships as GPT-6, a GPT-5-series release like GPT-5.7, or under another name. OpenAI itself simply calls it ‘our next major model.’
- Is this the same as Google's Project Astra?
- No. Google DeepMind’s Project Astra is a universal-assistant research project unveiled in May 2024. OpenAI’s Astra is an unreleased frontier model first reported in July 2026. Same name, unrelated projects.
§ 01The record as of August 29, 2026
| Item | Status | Source |
|---|---|---|
| Astra is OpenAI’s next major model | Official | OpenAI, August 1 |
| Ten math and theoretical-CS results by an internal version, formalized in Lean | Official | OpenAI, August 1 |
| Significant advances in agentic coding and cybersecurity | Official | OpenAI, August 7 |
| Critical cyber threshold cannot be ruled out; development slowed | Official | OpenAI, August 7 and 18 |
| Strictest security safeguards; a significant number of workloads paused | Official | OpenAI, August 18 |
| Trained to enable persistent agents | OpenAI-attributed reporting | WIRED, August 27 |
| Will not be supplied to Cursor | Official | OpenAI, August 28 |
| Multi-agent collaboration on long-running tasks | Reporting | The Information, July 31 |
| Demonstrated to policymakers in Washington | Reporting | The Information and The Washington Post, July 31 |
| Naming undecided: GPT-6, GPT-5.7, or other | Reporting | The Information, July 31 |
| Release date | None exists | - |
| Pricing, tiers, API name, model card | Nothing announced | - |
| “Launching week of August 10”, “mewfour”, parameter counts | Rumor, uncorroborated | social media |
§ 02The month Astra went from secret to slowed
The whole public life of this model fits in thirty days, and the sequence explains its strange status.
July 31. The Information breaks the story: OpenAI is preparing a new model family tentatively called Astra, built for long-running tasks with multiple agents collaborating, and Sam Altman has been demonstrating it to policymakers and regulators in Washington that week. The Washington Post independently reports the DC demos: multiple AI assistants dividing work, unsolved math problems, one person doing work normally spread across specialists. Notably, per The Information, OpenAI had not decided whether this ships as GPT-6, GPT-5.7, or something else.
August 1. OpenAI names it. In a research publication, the company credits “an internal version of Astra, our next major model” with ten results that resolve or substantially advance long-standing open problems - sphere packing, non-sofic groups, Connes’s rigidity conjecture, multicolor Ramsey numbers, and six more - each formalized afterward in Lean. The publication’s one cost detail: the tokens behind all ten results would run roughly $2,000 at GPT-5.6 Sol API rates.
August 7. The brakes. OpenAI announces that internal evaluations “over the past few days indicate significant advancements in agentic coding and cybersecurity,” strong enough that “we cannot rule out critical cyber capabilities under our Preparedness Framework” - a conclusion reached, per the post, the night before. Astra activities that do not meet strengthened security controls are paused; universal chain-of-thought monitoring is applied to every agentic use of Astra including training; government agencies and safety organizations are brought in to test. Axios reports OpenAI also told the US administration it planned to delay the release.
August 18. The scope becomes clear. In its pacing post, OpenAI discloses a two-week pause in frontier RL training, its largest planned frontier run still on hold, and - specific to Astra - that “a significant number of workloads remain paused” until they meet the new security bar. Astra and cyber workloads now carry the company’s strictest safeguards.
August 26. The Hugging Face technical report untangles the family tree. The incident’s primary driver was IM1, an internal-only research model. Astra “was not involved in exploiting Hugging Face” - OpenAI’s own words from August 7, reaffirmed by the report. But the report also notes a different model from the same family as Astra, with different post-training, took part in later infrastructure-compromise activity. Same lineage, different animal, and OpenAI is explicit that behavior does not transfer across family members.
August 27-28. The persistence thread and the first exclusion. WIRED reports OpenAI has trained forthcoming models, including Astra, to enable persistent agents - the context for the Codex Persistent mode discovery the same week. And OpenAI’s Cursor announcement names Astra as a model Cursor will never receive, citing “a new level of accountability to ensure our upcoming model, Astra, is being used in accordance with our terms.”
§ 03What Astra actually appears to be
Read the official statements and the credible reporting together and a consistent shape emerges: not a bigger chatbot, but a system built for long-horizon, multi-agent, high-autonomy work. The Information’s sources describe agents dividing and coordinating work over extended periods. OpenAI’s own math publication shows the mode of use: point it at a hard open problem, let it run, get a manuscript-grade argument back. The coding and cybersecurity gains that triggered the Critical review are the same capability measured on a different axis - an agent that can carry a long technical campaign to completion, whichever direction it is pointed.
That is also why the safety story and the capability story are the same story. OpenAI’s Critical threshold is not about what a model knows; it is about what a model can autonomously do against hardened systems, end to end, from a high-level goal. A model that can resolve Erdős problems unsupervised is impressive; the identical persistence pointed at infrastructure is what the Preparedness Framework exists to catch. OpenAI slowing its flagship for exactly this reason is the most concrete evidence yet that the industry’s next competition is not raw capability but the boundaries around it.
§ 04What we make of it
Our honest read, clearly labeled as our read: Astra is further confirmation that the frontier has moved from answering to working - long-horizon, self-directed, multi-agent. That is the bet the whole agent category is built on, and OpenAI training its next flagship “to enable persistent agents” is the strongest endorsement the category has received. The open question Astra cannot answer from inside a lab is the one every deployer faces: persistent capability is only useful once permissions, monitoring, and approvals are architectural. OpenAI’s own August is a month-long case study in that sentence.
§ 05The tracker
What we are watching, and will update here same day:
- A release date or launch window - none exists; any page claiming one ahead of OpenAI is guessing.
- The shipping name - GPT-6, a GPT-5-series release, or Astra itself.
- The Critical determination - OpenAI finishing its benchmarking: a formal classification either way, published scores, or third-party testing results.
- The paused workloads - OpenAI saying the security migration is complete, or its largest frontier RL run resuming.
- Availability decisions - which products and partners get Astra (Cursor officially will not), and any pricing signal.
- Persistence in product form - whether Astra ships with the persistent-agent capabilities OpenAI says it is trained for, and what Persistent mode becomes when it does.
§ 06Update log
This is a living page; when the story moves, the update lands here. As of August 29, 2026: officially named, officially unscheduled. OpenAI cannot rule out Critical cyber capability, a significant number of Astra workloads remain paused, and no date, price, or availability path exists. The most honest sentence available, so it is the one this page leads with.
Q1What has OpenAI officially confirmed about Astra?
Four things, each in an official OpenAI publication: it is the next major model (August 1); an internal version produced ten mathematical and theoretical-computer-science results, later formalized in Lean (August 1); evaluations showed advances in agentic coding and cybersecurity strong enough that OpenAI cannot rule out the Critical threshold (August 7); and Astra workloads now run under OpenAI’s strictest security safeguards, with a significant number paused (August 18). Everything else is reporting or rumor.
Q2Why is Astra delayed?
Cybersecurity. On the night of August 6, OpenAI concluded it could not rule out Astra meeting the Critical cyber threshold of its Preparedness Framework - the level reserved for capabilities that could create unprecedented new routes to severe harm. Critical-track models require safeguards during development, not just at release, so OpenAI paused every Astra workload that did not meet its upgraded security requirements. Axios reported OpenAI also told the US administration it planned to delay the release.
Q3What does the Critical cybersecurity threshold actually mean?
Per OpenAI’s August 7 definition, a model reaches Critical if it can either identify and develop working zero-day exploits across many hardened, real-world critical systems without human intervention, or devise and execute a novel end-to-end cyberattack against hardened targets given only a high-level goal. OpenAI says it cannot rule this out for Astra - which is a preliminary assessment, not a final classification, and no scores have been published.
Q4Did Astra hack Hugging Face?
No. OpenAI’s August 26 technical report attributes the incident primarily to an internal-only research model it calls IM1, which was never intended for release. The report separately notes a different model from the same family as Astra, with different post-training, was involved in later infrastructure-compromise activity. OpenAI states directly that Astra was not involved in the Hugging Face exploit.
Q5What will Astra cost?
Nothing is known. No API rate, subscription tier, or availability path has been announced. The only public number near Astra is OpenAI’s note that the tokens behind its ten math results would cost roughly $2,000 at GPT-5.6 Sol API rates - a research-compute estimate, not Astra pricing.
Q6Where will Astra be available?
Unannounced - but one exclusion is already official: OpenAI’s August 28 Cursor announcement says future models, including Astra, will not be supplied to Cursor following its acquisition by SpaceX. Access so far has been limited to OpenAI internal work, demonstrations to US policymakers in late July, and reported private previews for selected customers in early August.