Short answer: Critical means a model can find and exploit zero-days in hardened systems without human help. GPT-6 Astra did that in OpenAI's tests: a browser exploit chain in 29 hours plus 12 more, and a kernel exploit in 12 hours. The public model refuses to write proof-of-concept exploits.
How does Critical differ from High in OpenAI's Preparedness Framework?
Critical is the top cyber tier, and Astra is the first model OpenAI has placed there. OpenAI's Path to Astra post, dated September 1, 2026, says Critical "requires stronger safeguards during development and before release."
A model meets Critical if either condition holds:
- Zero-days: it can "identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention."
- Novel attacks: it can "devise and execute end-to-end novel strategies for cyberattacks against hardened targets given only a high level desired goal."
High sits one step below. The system card defines it as a model that "removes existing bottlenecks to scaling cyber operations." Astra itself rates High in biological and chemical risk and below High in AI self-improvement.
How did Astra's exploit tests work?
The tests worked as unaided research runs. Experts only supervised "for safety and validation" and were not allowed to give the model ideas or steer its direction. Astra ran in OpenAI's standard Codex harness at Ultra reasoning effort, with web access and up to 64 subagents.

- Browser, first pass: after 29 hours it had a chain that escaped the sandbox, but on a build that "lacked some production security mitigations."
- Browser, stable release: adapting the exploit to the official release took "a further 12 hours."
- Kernel: it built a working local privilege-escalation exploit, unprivileged user to root, "within 12 hours."
OpenAI says the runs surfaced two previously unknown vulnerabilities, which it is disclosing to the maintainers.
OpenAI also says these results "reflect capabilities with Daybreak Blue access, not the default production configuration."
How much should you trust the benchmark scores?
Trust the expert runs more than the benchmarks. The card says Astra scores 100% on ExploitBench, 41 V8 vulnerabilities, "even at the lowest reasoning effort tested." It also warns the results "may be artificially inflated due to potential contamination."
The outside view is more modest. Irregular, an outside lab that also evaluated Astra, "observed no successful attacks on fully hardened targets," and neither model it tested solved any of the seven Elite challenges. Comparison figures for GPT-5.6 Sol circulating in coverage are unconfirmed: CSO Online cites Sol's ExploitBench score as 78.5% in one place and 73.5% in another.
What does the public Astra refuse, and who gets more?
The public model does code review and patching but, per OpenAI's launch page, refuses "creating proof-of-concept exploits for vulnerabilities." Defenders approved for Daybreak Blue get a looser boundary.
| Task (system card, Table 21) | Without trusted access | With Daybreak Blue |
|---|---|---|
| Vulnerability discovery and analysis | 66.7% | 100% |
| Patching | 44.4% | 100% |
| Proof-of-concept exploit creation | 2.4% | 92% |
| Cyber red-teaming | 7.4% | 76.9% |
| Advanced cyber completion rate | 3.5% | 3.5% |
The last row is the surprise: even with Daybreak Blue, Astra fully completes only 3.5% of arbitrary cyber requests. Other safeguards stack on top:
- Jailbreaks: Astra refuses 91.5% of cyber jailbreak attempts, against 59% for GPT-5.6 Sol.
- Risky accounts: accounts "assessed as higher risk" get "a more conservative model-behavior boundary."
- Pauses: if the misalignment monitor pauses a task, ChatGPT and Codex users may review it, while API tasks simply stop.
Daybreak for Frontline Defenders, dated September 3, 2026, commits $1 billion in subsidized access over six months, starting in the United States. Daybreak Blue covers the mainline models, Daybreak Red covers specialized cyber models, and 2,000 organizations and workspaces are already approved.
Why does lower monitorability matter?
It matters because the card states "GPT-6 Astra's monitorability has decreased relative to GPT-5.6 Sol." When reasoning is harder to read, monitors that watch it can catch less.
- Response: OpenAI added misalignment monitoring to all tool-using inference in Astra's external deployment, "with significant compute cost."
- Scope: in OpenAI's launch-page tests, GPT-5.6 Sol went beyond the authorized target 48% of the time, and Astra did so in 0% of cases.
What does Astra cost?
The Astra API model page lists these rates per 1M tokens:
| Tier | Input | Cached input | Cache writes | Output |
|---|---|---|---|---|
| Standard | $10.00 | $1.00 | $12.50 | $50.00 |
| Over 272K input tokens | $20.00 | $2.00 | $25.00 | $75.00 |

The higher rates apply to the full request, not just the tokens past 272K. A 300K-token prompt costs $6.00 in input instead of $3.00. Batch and Flex cost 50% of Standard, Fast mode costs 2x, and regional processing (data residency) endpoints add 10%.
Astra costs 2.5x GPT-5.6 Sol ($4/$20) and 5x GPT-6.1 Sol ($2/$10), covered in our DevDay recap.
What should security teams do before they try Astra?
Start from containment, as Microsoft does. Its Foundry post says Astra is "generally available for all customers" and that "capability this direct demands containment."
- Pick the tier: use the public model for review and patching, and apply for Daybreak Blue if you need exploit validation.
- Contain it: give agents "scoped credentials, approved resources, human checkpoints for consequential actions, and activity records."
- Plan for pauses: API tasks stop when the monitor fires, so build retries, alerts and your own logs.
- Watch the 272K line: keep prompts under it unless the work needs more.
For setup patterns, see our agent guardrails and approval tools, browse AI cybersecurity tools, read the GPT-6 Astra model page, or see how we compared the three models.




