OpenAI released GPT-6 Cyber (`gpt-6-cyber`) on 5 November 2026 via the API and Codex. It is a cyber-specialised variant of GPT-6 Astra, built for security work — finding vulnerabilities and testing defences.
The release carries a "Critical" rating under OpenAI's Preparedness Framework — one step up from the "High" rating given to GPT-5.6 Cyber, the previous model in the line. That rating is why access is gated: GPT-6 Cyber is available only to Daybreak Red customers, and even they need a separate model approval. GPT-5.6 Cyber opened to all Daybreak Red customers on the same day.
OpenAI says the model found 81 previously unknown high-severity flaws in Firefox and produced working exploits for 948 Linux kernel vulnerabilities. It reports an ExploitGym score of 48.5%. All of these figures are OpenAI's own self-reported results — they have not been independently verified, and the 48.5% score should be read as the vendor's claim, not as an independent measurement.
Key facts
| Item | Details |
|---|---|
| Model | GPT-6 Cyber (`gpt-6-cyber`), cyber-specialised variant of GPT-6 Astra |
| Released | 5 November 2026, via API and Codex |
| Preparedness Framework rating | "Critical" (up from "High" for GPT-5.6 Cyber) |
| Access | Daybreak Red customers only, plus separate model approval |
| ExploitGym | 48.5% — OpenAI's own self-reported figure, not independently verified |
| Reported findings | 81 previously unknown high-severity Firefox flaws; working exploits for 948 Linux kernel vulnerabilities |
| API pricing | $20 input / $100 output per 1M tokens; $2 cached input |
| GPT-5.6 Cyber | Opened to all Daybreak Red customers the same day |
What this means
A "Critical" rating is the strongest signal OpenAI's framework gives about a model's cyber capability — and the gated release is the other side of that coin. This is not a model any developer can call from the API: access runs through Daybreak Red, OpenAI's programme for vetted security researchers and organisations, with a separate approval step for this model.
For security teams, the headline claims — 81 new Firefox flaws, working exploits for 948 kernel vulnerabilities — show what the model is aimed at: offensive security research done at machine speed. For everyone else, the practical point is narrower: GPT-6 Cyber is a specialised tool with a controlled rollout, and its benchmark numbers are OpenAI's own until independent testing catches up.
Related: our post on OpenAI's kernel pipeline covers the 948 drafted patches that pipeline produced — a separate item from this model launch.
A note on verification
The release date, rating, access terms and pricing are as reported for the 5 November 2026 release, re-verified against 9 independent sources in the 21 November 2026 re-verify pass. The ExploitGym score and the flaw/exploit counts are OpenAI's own self-reported figures; no independent verification of them exists yet.
FAQs
What is GPT-6 Cyber?
A cyber-specialised variant of GPT-6 Astra, released by OpenAI on 5 November 2026 via the API and Codex, built for security work such as finding vulnerabilities.
Why is it gated?
It carries a "Critical" rating under OpenAI's Preparedness Framework — up from "High" for GPT-5.6 Cyber. Access is limited to Daybreak Red customers with a separate model approval.
What does it cost?
$20 per million input tokens and $100 per million output tokens, with cached input at $2 per million tokens.
Is the 48.5% ExploitGym score independently verified?
No. It is OpenAI's own self-reported figure, and the flaw and exploit counts are likewise the vendor's claims until independent testing confirms them.
Source: OpenAI release, 5 November 2026 (API and Codex); re-verified 21 November 2026. Last checked: 21 November 2026.
Comments (0)
Share your question or experience about this syllabus. Your comment will appear after review.
No comments yet — be the first to share your thoughts!