GPT-6 Cyber

GPT-6 Cyber is an AI model for investigating software vulnerabilities, reproducing them in authorised test environments and checking fixes. OpenAI released it on 5 November 2026. Access is restricted through Daybreak Red, with eligibility and permission requirements.

Access requires approval

The Daybreak guide describes access for approved security work. Developers should check their organisation’s eligibility and the tools and targets permitted by the programme. OpenAI’s model documentation lists the API identifier gpt-6-cyber, with Responses and Chat Completions support.

The model reads text and images

The model page lists a 1,050,000-token context, with up to 922,000 input tokens and 128,000 output tokens. It accepts text and images and produces text. Tokens are the units used to measure the content processed by the model.

API usage has separate input and output charges

Standard short-context prices checked on 9 November are $20 per million input tokens, $2 for cached input and $100 for output. Cache writes and long-context requests have separate rates in the pricing documentation. The total bill depends on the request size, generated output and caching behaviour.

Security scores describe particular evaluations

OpenAI’s safety report reports 87.02% pass@1 on SEC-Bench Pro, 48.45% intended-exploit success on ExploitGym and 36.94% on its Hardened variant. The hardened tasks add protections such as address randomisation and sandboxing where supported.

These are vendor-reported measurements. Scores from different task versions, tool access and retry budgets should be read separately. Our model security capability tracker records the individual metrics and their source conditions.