GPT-5.6-Cyber

// Overview

GPT-5.6-Cyber is OpenAI's purpose-trained cybersecurity model, announced on August 10, 2026 alongside a restructuring of the Daybreak program into two access tiers. It is built on GPT-5.6 Sol and trained to improve on specialized security tasks — finding zero-day vulnerabilities, developing exploit chains, validating exploits — while refusing far less often on higher-risk dual-use cyber work than the general-purpose model.

Access runs through Daybreak Red, the tier for security researchers doing vulnerability research, exploit validation and penetration testing; Daybreak Blue is the broader tier, giving GPT-5.6 Sol with tailored safeguards for defensive work such as vulnerability discovery, secure code review, malware analysis, incident response and patch validation. Entry to either tier requires identity verification, account security measures, monitoring and legal declarations, and hardware security keys become mandatory for all Daybreak accounts on September 1, 2026.

The model takes text and image input and returns text, with a 400,000-token context window, up to 128,000 output tokens, and a February 16, 2026 knowledge cutoff. OpenAI prices it at $12.50 per million input tokens, $1.25 per million cached input tokens and $75.00 per million output tokens, and notes that requests over 272K input tokens are billed at 2× input and 1.5× output for the whole request. Only the Responses API endpoint is supported. OpenAI reports using the model in real-world vulnerability research, including work that surfaced previously unknown flaws in Chrome's V8 engine.

// Benchmarks

This model's scores

1. Advanced cybersecurity request completion (OpenAI internal set)95%

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

// Pricing

Prompts over 272K input tokens are priced at 2× input and 1.5× output for the full request. Access is granted through the Daybreak Red application, not the public price list.

Pricing source ↗

// Strengths

• Answers the authorized security questions a general-purpose model blocks — 95.0% completion on OpenAI's advanced cybersecurity request set, against 1.5% for GPT-5.6 Sol
• Trained for zero-day discovery, exploit-chain development and exploit validation rather than defensive triage alone
• Outperforms both GPT-5.6 Sol and GPT-5.5-Cyber on ExploitGym, which tests turning known vulnerabilities into working exploits
• 400K-token context with 128K max output — enough for whole-codebase review and long agentic security runs
• Inherits GPT-5.6 Sol's frontier reasoning, vision input and tool use

// Best for

• Reach for it when an approved research team needs a model that will engage with authorized vulnerability research, exploit validation and penetration testing.
• Reach for it for zero-day hunting and exploit-chain development in a monitored, sandboxed environment under coordinated disclosure.
• Use Daybreak Blue with GPT-5.6 Sol instead for the defensive majority of security work — vulnerability discovery, secure code review, malware analysis, incident response and patch validation.
• Use standard GPT-5.6 Sol for everything outside authorized security work: it is a fifth of the input price and carries the full 1.05M-token context.

// 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.