Quick answer
On August 7, 2026, OpenAI disclosed that preliminary internal evaluations of Astra, an upcoming model, showed enough cybersecurity capability that the company could not rule out the Critical threshold in its Preparedness Framework.
That wording matters. It is a safety and governance disclosure, not a launch announcement, a final Critical classification, or proof of public access. The announcement provides no Astra model ID, price, API endpoint, release date, or Flowith integration.
Use the Astra versus GPT-5.6 Sol comparison for the threshold distinction, the Critical cyber controls guide for the operating controls, and the Astra cyber readiness checker for a bounded evaluation gate.
What OpenAI actually disclosed
OpenAI says recent internal benchmarks and expert assessments indicated significant progress in agentic coding and cybersecurity. While evaluation continued, the evidence was strong enough that the company could not rule out Critical cyber capability.
Under the framework definition summarized in the announcement, Critical covers autonomous ability to find and develop functional zero-day exploits across many hardened critical systems, or to plan and execute novel end-to-end attacks against hardened targets from a high-level goal. That is a capability threshold, not a score readers can reproduce from the announcement.
OpenAI also explicitly says Astra is upcoming and was not involved in the Hugging Face exploitation incident. Do not connect the model to an incident, endpoint, product surface, or model family without a first-party statement.
Preliminary is not confirmed
Keep four states separate:
| State | What it establishes |
|---|---|
| Preliminary internal signal | Current tests are concerning enough to justify stronger safeguards and more evaluation |
| Confirmed framework assessment | A completed assessment reaches a documented threshold |
| Deployment decision | OpenAI has decided how, where, and to whom a model may be made available |
| Public product release | A named product, model ID, access path, terms, and supported surfaces are documented |
The August 7 post establishes only the first state and the resulting control response. It does not establish the other three.
Controls named in the announcement
OpenAI says it scaled robustness testing and introduced or strengthened:
- isolated testing environments;
- restricted network and tool access;
- stronger model-weight protection and encryption;
- additional monitoring and detection;
- sandboxed execution;
- pauses on internal Astra activity that did not meet the stronger controls;
- monitoring of risky actions and misalignment across agentic Astra applications;
- testing with government agencies, safety organizations, and selected external partners.
These are provider-level measures. An evaluator still needs its own written authorization, environment isolation, least-privilege tools, approval boundaries, audit logs, incident handling, and accountable human review.
Do not infer an Astra product identity
OpenAI’s current developer documentation separately lists gpt-5.6-cyber for approved defensive cybersecurity work. The Astra announcement does not say Astra became that model, that Astra is available through Daybreak, or that any current alias resolves to Astra.
The safe rule is simple: treat Astra as the upcoming model named in the disclosure and treat documented API models as separate products unless OpenAI explicitly links them.
Frequently asked questions
Has OpenAI released Astra?
No release is established by the August 7 disclosure. OpenAI calls Astra an upcoming model and does not provide a public model ID, access path, price, or launch date there.
Did OpenAI confirm Astra is at the Critical cyber threshold?
No. OpenAI says preliminary internal evaluations were strong enough that it could not rule out Critical capability while benchmarking and expert assessment continued.
Is GPT-5.6 Cyber the public name for Astra?
The reviewed official sources do not establish that mapping. Current developer documentation lists gpt-5.6-cyber separately, so do not treat it as Astra without an explicit OpenAI source.
Official sources
- OpenAI: Responding to the next frontier of critical cyber capabilities
- OpenAI Preparedness Framework
- OpenAI API: GPT-5.6 Cyber
Source check: August 12, 2026. Recheck the announcement, framework, model documentation, and access program before making a deployment decision.