GPT-5.6-Cyber
2026-08-10 · Publisher : OpenAI · Domain : Artificial Intelligence
OpenAI launches GPT-5.6-Cyber, a specialized model built on GPT-5.6 Sol, and opens two Daybreak access tiers: Blue for everyday defense, Red for exploitation and advanced research.
What changes
- Trained to find zero-day flaws and develop exploit chains, with fewer refusals on dual-use tasks.
- Daybreak Blue gives access to GPT-5.6 Sol without the system safeguards, Daybreak Red to GPT-5.6-Cyber.
- Used by OpenAI to discover two previously unknown vulnerabilities in V8, fixed by Google (CVE-2026-15903).
- Rated “High” and not “Critical” in the Preparedness Framework; access subject to identity verification and legal attestations.
- Subsequently available on Amazon Bedrock for Daybreak customers.
### Specifications - Context: 400,000 tokens, maximum output 128,000 tokens, text and image input. - Knowledge cutoff: February 16, 2026. - Availability: Responses API, upon approval to the Daybreak program (alias gpt-daybreak-red-latest). - API identifier: gpt-5.6-cyber. ### Pricing | Input | Cached input | Cache write | Output | | --- | --- | --- | --- | | $12.50 | $1.25 | $15.625 | $75.00 | Prices per million tokens (USD). No long-context tier published. ### Benchmarks | Completion rate on advanced cyber requests (internal measure) | Result | | --- | --- | | GPT-5.6-Cyber (Daybreak Red) | 95.0% | | GPT-5.5-Cyber | 57.3% | | GPT-5.6 Sol with Daybreak Blue | 2.0% | | GPT-5.6 Sol with standard safeguards | 1.5% | Scores published by the vendor. The post also claims that GPT-5.6-Cyber outperforms Sol and GPT-5.5-Cyber on ExploitGym, with no figure given in the text. ### Key takeaways Offensive capabilities have moved out of the general catalog and go through a controlled access program. Security teams must sign up for Daybreak before any testing.
Links
Tags : LLM, OpenAI, GPT-5, Cybersécurité, Daybreak, Accès restreint