GPT-5.6-Cyber: OpenAI's Daybreak Red Cybersecurity Model

GPT-5.6-Cyber is OpenAI's answer to a problem every frontier lab now has: general models refuse the dual-use prompts that professional defenders need, while attackers race to use the same class of tools. The interesting story is the access model plus an uneven specialty vs Sol, not a blanket Elo win.

At a Glance

• Release date: August 10, 2026
• Model ID: gpt-5.6-cyber (alias daybreak-red-latest)
• Organization: OpenAI
• Base: GPT-5.6 Sol (purpose-trained specialization)
• Access: Daybreak Red only (approved partners / orgs; identity verification; monitoring; hardware security keys for individual accounts from Sept 1, 2026)
• Pricing (API list): $12.50 input / $1.25 cached input / $75 output per 1M tokens
• Context: 400K total; max input 272K; max output 128K. Prompts with >272K input billed 2x input / 1.5x output
• Modalities: Text + image in, text out; Responses API (Chat Completions not supported per markdown docs)
• Aliases: daybreak-red-latest → gpt-5.6-cyber; daybreak-blue-latest → gpt-5.6-sol
• Preparedness: Cyber capability rated High, below Critical (OpenAI, self-reported)

What's Actually New

1. Daybreak splits into Blue and Red

OpenAI expanded Daybreak into two tiers:

• Daybreak Blue: Frontier general models (including GPT-5.6 Sol) with safeguards calibrated for broad defensive work. Alias: daybreak-blue-latest.
• Daybreak Red: Purpose-trained cybersecurity models, including GPT-5.6-Cyber, for authorized vulnerability research, exploit validation, and security testing. Alias: daybreak-red-latest.

2. Same Sol base, uneven specialty

OpenAI says GPT-5.6-Cyber is built on GPT-5.6 Sol and trained for specialized dual-use cyber tasks with fewer refusals. Specialty is uneven vs Sol (self-reported / OpenAI docs):

• Beats Sol on ExploitGym and internal zero-day work
• Lags Sol on vulnerability discovery and report writing
• Sol is more efficient on ExploitBench 300-turn

In an internal OpenAI test covered by press (exploit chains, auth bypass, privilege escalation), GPT-5.6-Cyber answered about 95% of requests versus roughly 1.5% for Sol under standard safeguards and ~2% under Daybreak Blue. Treat those figures as self-reported / press-relayed, not an independent LLM Stats verification.

3. Access is still the product

Daybreak Blue and Red require approval for authorized defensive work. Controls include identity verification, legal attestations, monitoring, and (for individuals) hardware security keys starting September 1, 2026. Apply via https://openai.com/daybreak/partners

OpenAI cites real vulnerability research including CVE-2026-15903 (Chrome V8). That CVE ID is OpenAI's claim; we are not independently verifying the advisory here.

Benchmarks

Do not treat Sol's general leaderboard scores as Cyber scores.

Pricing & Availability

When to Use It

Reach for Daybreak Blue / Sol first for defensive workflows Sol can already handle under Blue safeguards, and for vuln discovery / report writing where Sol leads.

Request Daybreak Red / GPT-5.6-Cyber when authorized scope needs exploit validation, ExploitGym-class work, or red teaming on systems you own or have written permission to test.

Do not treat this as a general coding or agent upgrade. It is a gated specialist with a narrower context window and higher list price than Sol.

Outlook

GPT-5.6-Cyber is both a distribution event and a priced specialist SKU. OpenAI split Blue vs Red, published API rates above Sol, and documented an uneven win/loss pattern vs Sol on cyber tasks. Until independent evals exist, treat ExploitGym / refusal-gap numbers as vendor or press-relayed signals, not a universal ranking.

Sources: https://developers.openai.com/api/docs/models/gpt-5.6-cyber ; OpenAI Daybreak announcement; press-relayed internal test numbers only where labeled.