GPT-6 Cyber

// Overview

GPT-6 Cyber is OpenAI's cybersecurity model of the GPT-6 generation, released on November 5, 2026. It is a cyber-specialized variant of GPT-6 Astra for authorized security work — vulnerability discovery, exploit development and web pentesting — and OpenAI's system card describes substantially stronger cyber capability and better alignment than GPT-5.6-Cyber. On OpenAI's evaluations it completes 96% of exploit-development requests and 97% of red-teaming requests without refusing.

Access runs through Daybreak Red, OpenAI's tier for advanced, authorized security work. The model starts with a limited group of approved customers through the API and Codex, and OpenAI says it will widen to all Daybreak Red customers over the coming weeks. Organizations must use single sign-on, multi-factor authentication and API IP allowlisting; individuals go through identity and liveness checks, a professional suitability review and a legal attestation, and must use at least one physical FIDO2 security key.

Under OpenAI's Preparedness Framework, GPT-6 Cyber is rated Critical for cybersecurity (the same as GPT-6 Astra), High but not Critical for biological and chemical risk, and below High for AI self-improvement. Extra safeguards target high-risk help with industrial controls, essential services and election systems, developed with Dragos, and a conservative biological safety mode is on for every user. The API lists a 1,050,000-token context window, up to 128,000 output tokens, text and image input, and a price of $20 per million input tokens and $100 per million output tokens.

// Benchmarks

This model's scores

1. ExploitGym (intended-exploit success)48.45%
2. ExploitGym Hardened36.94%
3. SEC-Bench Pro (pass@1)87.02%
4. SRE-Bench (pass@1)92.84%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Requires Daybreak Red access and organization approval.

Pricing source ↗

// Strengths

• Beats GPT-6 Astra on OpenAI's exploit benchmarks — 48.45% vs 42.43% on ExploitGym and 36.94% vs 29.69% on ExploitGym Hardened
• Higher pass@1 than Astra on SEC-Bench Pro (87.02% vs 85.38%) and SRE-Bench (92.84% vs 87.98%)
• Stays in scope: on the ExploitGym honeypot it went after the out-of-scope target in 0.07% of runs without safety instructions, against 5.51% for GPT-5.6-Cyber
• Irregular's external tests found it solved all 11 CyScenarioBench challenges at least once, against 8 for GPT-6 Astra
• 1.05M-token context window with 128K max output for long agentic security runs

// Best for

• Reach for it when an approved Daybreak Red team needs to investigate a vulnerability, reproduce it in an authorized test environment, or verify a fix.
• Reach for it for long exploit-development and pentest runs in Codex, with sandboxing and Auto-Review, which OpenAI recommends for this model.
• Use Daybreak Blue instead for most defensive work; OpenAI points most defenders there and is widening an alpha of GPT-6 Astra in that tier.
• Keep an eye on its blocker reports: 20.29% of its blocker-reporting responses were flagged in OpenAI's tests, against 11.66% for GPT-6 Astra.

// How to access

// GPT Cyber — every version

The full lineage of the GPT Cyber line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.