Unreleased / no neutral benchmark yet
OpenAI Astra: The model OpenAI is keeping close.
Astra has generated serious mathematics, triggered a Critical cyber review, and become the model OpenAI named while withdrawing future releases from Cursor. Its capability and distribution are now one story.
Astra is already a research result, a security problem, and a distribution decision. It remains impossible for anyone outside OpenAI to measure independently.
What is established
Three facts changed the story.
Public evidence still comes mainly from OpenAI. These are company claims and policy decisions, with independent benchmarks still absent.
Ten research results
OpenAI says an internal Astra version resolved or substantially advanced ten open problems in mathematics and theoretical computer science. The model later formalized the arguments in Lean.
Inspect source ->A Critical cyber review
OpenAI said it could not rule out its Critical cybersecurity threshold: autonomous zero-day exploitation across hardened systems or novel end-to-end attacks from a high-level goal. Astra was not involved in the Hugging Face breach.
Inspect source ->Future models leave Cursor
After SpaceX acquired Cursor, OpenAI said it would stop supplying future models. It explicitly cited responsibility for keeping Astra use within its terms.
Inspect source ->The Cursor decision
Astra will not ship through Cursor.
Cursor became part of SpaceX on August 14. Two weeks later, OpenAI proposed ending its model contract on November 12 and said future models would not be provided to Cursor.
OpenAI's stated reason is trust: it says prior conduct by Musk companies makes contract and terms-of-service compliance too uncertain for Astra. Cursor says SpaceX gives it the compute to build stronger, cheaper models of its own.
The immediate result is simple. Developers who want Astra will need an OpenAI-controlled or OpenAI-approved channel.
Daybreak is the clue
The model is only half the product.
OpenAI already routes cyber capability through different access tiers. Daybreak shows how one model family changes when the allowed work, safeguards, and customer verification change.
Daybreak Blue and Red are access tiers. Blue exposes GPT-5.6 Sol; Red exposes purpose-trained cyber models such as GPT-5.6-Cyber.
Daybreak Blue
DefenseFrontier general-purpose models, including GPT-5.6 Sol, configured for approved defensive security work.
Vulnerability discovery, secure code review, malware analysis, incident response, and patch validation. Higher-risk requests can still be refused.
Daybreak Red
ResearchPurpose-trained cyber models, including GPT-5.6-Cyber, for vetted vulnerability research, exploit validation, and red teaming.
OpenAI reports a 95% completion rate on its advanced cyber-request eval. Blue scored 2%. Sol/Blue still led the standard 300-turn ExploitBench setting and used fewer tokens.
Those results measure permission and specialization. They do not establish a universal model ranking. Astra may use a similar access model; OpenAI has not announced one.
Evidence boundary
What we still do not know
Astra has no public neutral score. It does not belong on a model leaderboard yet.
- 01Release date and public model ID
- 02ChatGPT, Codex, API, cloud, or gated availability
- 03Price, context window, rate limits, and cost per task
- 04Independent benchmark results and production reliability
- 05The final system card and deployment safeguards
Astra turns the operating boundary into part of the product.
The hard question is who can run the model, with which tools, under whose supervision, and who bears the cost when it finds an unexpected route.
Short answers
OpenAI Astra FAQ
What is OpenAI Astra?
OpenAI describes Astra as its next major model. The company says an internal version produced ten mathematics and theoretical-computer-science advances and later showed enough cyber capability that OpenAI could not rule out its Critical threshold.
Will OpenAI Astra be available in Cursor?
OpenAI said on August 28, 2026 that it will not provide future models to Cursor after Cursor's acquisition by SpaceX. OpenAI explicitly named Astra while explaining that decision.
Is OpenAI Astra publicly released?
No public Astra model, API price, system card, or independent benchmark result was available as of August 30, 2026. Astra therefore cannot yet be ranked against public frontier models with neutral evidence.