GPT-6 Cyber explained: Daybreak Red, price, scores.

On 5 November 2026, OpenAI introduced GPT-6 Cyber, a cyber-specialized variant of GPT-6 Astra for authorized security work. OpenAI says the model is better at finding and validating exploitable vulnerabilities than GPT-5.6 Cyber, and far more reliable at staying inside the scope researchers set. Access starts through Daybreak Red for approved customers. Here is what OpenAI’s launch post, system card, model page and Daybreak docs confirm — and how GPT-6 Cyber sits on The Model Atlas.

What is GPT-6 Cyber?

OpenAI’s system card↗ (opens in a new tab) describes GPT-6 Cyber as a cyber-specialized variant of GPT-6 Astra with extra post-training for cybersecurity capabilities and fewer refusals on authorized security requests. The goal OpenAI states is to expand the security work the model can do while keeping it inside the user’s authorized scope and still refusing work outside cybersecurity.

The API model ID is `gpt-6-cyber`. OpenAI’s model page↗ (opens in a new tab) lists text and image inputs, text output, a 1,050,000-token context window, up to 922,000 input tokens, up to 128,000 output tokens, and reasoning-token support. The API changelog↗ (opens in a new tab) dated 5 November 2026 says to use it to investigate vulnerabilities, reproduce them in an authorized test environment, and verify fixes.

What’s new in GPT-6 Cyber

OpenAI positions GPT-6 Cyber as a major step up from GPT-5.6 Cyber and says it outperforms GPT-6 Astra on most cyber tasks, even when Astra does not refuse the request. Under OpenAI’s Preparedness Framework, OpenAI classifies GPT-6 Cyber as Critical in cybersecurity (up from High for GPT-5.6 Cyber), High in biological and chemical capabilities, and below High in AI self-improvement — the same designations OpenAI gives GPT-6 Astra.

Benchmarks OpenAI reports

• ExploitGym: OpenAI says GPT-6 Cyber scores 48.5%, up from 32.1% for GPT-5.6 Cyber (about 16 points, or a 51% relative gain). The system card lists 48.45% intended-exploit success versus 42.43% for GPT-6 Astra.
• Web penetration testing (no source code): OpenAI says 40.3% vulnerability recall versus 32.0% for GPT-6 Astra.
• RemoteExploitBench: OpenAI says GPT-6 Astra with Daybreak Blue refused 100% of tasks; GPT-6 Cyber completed 95%.
• ExploitGym Honeypot (stay in scope): without deployment prompts, OpenAI says Cyber avoided out-of-scope exploits in 99.93% of cases (up from 94.49% for GPT-5.6 Cyber); with prompts, 100%.
• Other system-card scores: ExploitGym Hardened 36.94% (Astra 29.69%); SEC-Bench Pro pass@1 87.02% (Astra 85.38%); SRE-Bench 92.84% (Astra 87.98%). On ExpertSWE general coding, OpenAI says Cyber 82.95% versus Astra 83.24%.

Real-world findings OpenAI cites

OpenAI says its researchers used GPT-6 Cyber to help secure real systems: 81 previously unknown high-severity Firefox vulnerabilities (with Mozilla releasing fixes for many), working exploits for 948 of 1,724 reproduced Linux kernel issues (versus 166 for GPT-5.6 Sol with the same harness), and 304 critical Linux disclosures with candidate patches. OpenAI also says teams at Cisco, Calif, CodeRabbit and others are using it for research and code review, including hardening OpenAI’s own training infrastructure.

Who GPT-6 Cyber is for

OpenAI frames GPT-6 Cyber for authorized defensive cybersecurity: vulnerability discovery, validating impact, developing fixes, long-running exploit development inside approved rules of engagement, and security review of large codebases. Daybreak Red is the gated program for that deeper testing. Daybreak Blue remains the path OpenAI describes for defensive tasks such as code review and incident work with mainline models like GPT-6 Astra and fewer unnecessary refusals.

OpenAI is also expanding Daybreak Blue alpha for GPT-6 Astra, making GPT-5.6 Cyber available to all Daybreak Red customers, and introducing Daybreak Red for Critical Infrastructure for approved defenders who need authorized testing of critical systems.

How to try GPT-6 Cyber

1. Confirm your organization has Daybreak Red access and model approval. OpenAI’s model page says your organization needs Daybreak Red access and approval to use GPT-6-Cyber; the account team can confirm access.
2. Apply or expand access through OpenAI’s Daybreak↗ (opens in a new tab) program links on the launch post. OpenAI says GPT-6 Cyber is initially available through the API and Codex to a select group of approved Daybreak Red customers, with progressive expansion to all Daybreak Red customers.
3. In the Responses API, set `model` to `gpt-6-cyber` and `access_programs.cyber` to `daybreak_red`, and include a stable `safety_identifier` as in OpenAI’s Daybreak guide↗ (opens in a new tab).
4. Follow OpenAI’s sandboxing, Auto-Review and human-oversight guidance for high-risk workflows. Daybreak Red API access requires IP allowlisting by default; Zero Data Retention customers generally must use Private Safety Processing.
5. Read the GPT-6 Cyber System Card↗ (opens in a new tab) before production use.

API pricing OpenAI lists

OpenAI’s GPT-6-Cyber model page and pricing table list these short-context text-token rates for `gpt-6-cyber`:

• $20 per 1M input tokens
• $2 per 1M cached input tokens
• $25 per 1M cache writes (30-minute cache; $50 per 1M for 12-hour cache writes)
• $100 per 1M output tokens

OpenAI’s pricing page lists those same Cyber-model short-context rates. Check the live pricing page for any later changes.

What people are building with it

GPT-6 Cyber is initially available through the API and Codex to selected approved Daybreak Red customers. OpenAI’s launch post highlights vulnerability research, code review and authorized exploit development. The post also includes quotes from Firefox, Calif, Cognition, Palo Alto Networks, Cisco, SentinelOne, CrowdStrike, Perplexity and CodeRabbit. The Model Atlas lists GPT-6 Cyber beside GPT-6 Astra; eligible creator video builds that name these models may be added to those model pages and /builds/ after curation.

Safety and access notes

OpenAI pairs the stronger cyber capabilities with stronger safeguards. The launch post and system card describe cyber-specific misalignment monitoring, Private Safety Processing for many Zero Data Retention customers, default IP allowlisting for Daybreak Red API access, credential-compromise monitoring, and new checks meant to block high-risk assistance involving industrial controls, essential services and election systems. OpenAI also says biological safety mode is enabled for every GPT-6 Cyber user. Access still requires identity verification, use restrictions, SSO/MFA (for organizations) and related Daybreak Red security requirements.