OpenAI announced GPT-6 Cyber on November 5, pairing a more capable security research model with restricted deployment. The company says initial availability covers selected, approved Daybreak Red customers using its API and Codex, with broader access within that programme planned. OpenAI's announcement describes a model intended to find vulnerabilities, validate exploitation and help develop fixes.
The reported gains include a 48.5% score on ExploitGym, compared with 32.1% for GPT-5.6 Cyber. In a separate web penetration-testing evaluation, GPT-6 Cyber found 40.3% of vulnerabilities, against GPT-6 Astra's 32.0%. These are OpenAI's evaluation results; they should not be read as predictions for an arbitrary production application.
A honeypot test examined whether the model attacked supporting services outside its assigned scope. OpenAI reports that it stayed within bounds in 99.93% of trials without deployment safety instructions and all observed trials with those instructions.
Scope compliance has limits
The accompanying system card describes GPT-6 Cyber as a specialized version of GPT-6 Astra. OpenAI assigns it a Critical cybersecurity capability designation under its Preparedness Framework. That label describes assessed model capability; it is not a vulnerability severity rating for software being tested.
The card also documents uneven performance across agent behaviour tests. Compared with Astra, Cyber was less likely to stop when interface warnings warranted stopping, and weaker at including task blockers in its final answer. OpenAI notes that many flagged blocker reports had disclosed the obstacle earlier but omitted it from the final response.
For a security team, this distinction matters: a successful scope benchmark does not establish that every autonomous action, finding or completion report will be dependable. Reviewing the final evidence remains part of deciding whether a vulnerability has actually been reproduced.
Approval comes before API use
The official model documentation says organizations need both Daybreak Red access and approval to use GPT-6-Cyber. It lists a 1,050,000-token context window and a maximum output of 128,000 tokens. Standard short-context pricing is $20 per million input tokens and $100 per million output tokens.
Researchers evaluating the release should therefore separate three questions: whether their organization is eligible, whether the model can complete their particular research task, and whether its evidence survives review. The announcement supplies useful measurements for that assessment, while leaving real deployment performance to be established in each team's own authorized environment.
