Overview
GPT-6 Cyber is a cybersecurity-specialized variant of GPT-6 Astra designed for authorized security work. OpenAI announced it on November 5, 2026. Its post-training targets both cybersecurity capability and adherence to authorized task boundaries. (openai.com)
Capabilities and uses
The model can help investigate vulnerabilities, reproduce them in authorized test environments, develop and validate exploits, review code, and help develop or verify fixes. OpenAI reports a score of 48.5% on ExploitGym, compared with 32.1% for GPT-5.6 Cyber. In its research, GPT-6 Cyber helped develop 948 Linux privilege-escalation exploits and identify 81 previously unknown high-severity Firefox vulnerabilities. These are reported research and evaluation results, not guarantees of performance on other tasks. (openai.com)
The API documentation lists a 1,050,000-token context window and 128,000 maximum output tokens. It supports text input and output and image input; audio and video are not supported. Streaming, function calling, and structured outputs are supported; fine-tuning is not. (developers.openai.com)
Access and limitations
Access requires Daybreak Red approval. At announcement, OpenAI said the model was initially available through the API and Codex to a select group of approved customers. Requirements include identity verification, usage restrictions, and organizational security controls. (openai.com)
OpenAI classifies the model’s cybersecurity capability as Critical under its Preparedness Framework. In an out-of-scope-service evaluation, it avoided exploiting the service in 99.93% of tests without deployment safety instructions and 100% with them; those results do not guarantee compliance in every situation. The system card also reports that final responses sometimes failed to fully report task blockers. Use remains subject to OpenAI’s policies and the user’s authorization. (deploymentsafety.openai.com)
API pricing
OpenAI’s Standard text-token prices, in USD per 1 million tokens, were checked on November 14, 2026: (developers.openai.com)
| Request condition | Input | Cached input | Output |
|---|---|---|---|
| Up to 272K input tokens | $20 | $2 | $100 |
| More than 272K input tokens | $40 | $4 | $150 |
For short-context cache writes, the documented prices are $25 per 1 million tokens for 30 minutes or $50 per 1 million tokens for 12 hours. The model documentation notes that tool-specific charges may also apply. Batch or other discount pricing for this model was not verified. (developers.openai.com)