GPT-6 Cyber explained: gated, $20/$100 and rated Critical

By LLMTest Lab · Nov 8, 2026 · 3 min read releasellmsecurity
On this page

On this page

  1. What is new
  2. What OpenAI reports
  3. How it compares on paper
  4. Who should care

Not tested by us yet. GPT-6 Cyber is available only to selected, approved customers of OpenAI's Daybreak Red program, and we are not one of them. Everything below comes from VentureBeat's launch report and other independent coverage of OpenAI's announcement, which we could not open directly. We'll update this post if we get access or an open endpoint.

GPT-6 Cyber is OpenAI's model for security work, built to find and exploit software vulnerabilities. It costs $20 per million input tokens and $100 per million output tokens, the most expensive cyber model in the comparison below. Unless you run a vetted security team, you cannot use it yet, and for everyone else the practical question is what a general model can do instead.

GPT-6 Cyber
Maker OpenAI
Announced 2026-11-05
Input price $20 per M tokens
Output price $100 per M tokens
Cached input $2 per M tokens
Access Selected, approved Daybreak Red customers (API and Codex)
Risk rating "Critical" for cybersecurity in OpenAI's Preparedness Framework

What is new

The previous model, GPT-5.6 Cyber, was rated "High" on OpenAI's cyber scale. GPT-6 Cyber moves to "Critical", the same tier OpenAI gave GPT-6 Astra. Per VentureBeat, input is about 60% more expensive than GPT-5.6 Cyber ($12.50) and output about 33% more ($75).

Access is stricter than for a normal API model. Applicants upload information for review, API use generally requires approved network addresses, and zero-data-retention customers generally must use Private Safety Processing, which keeps encrypted interaction records for 30 days. OpenAI also introduced a Daybreak Red for Critical Infrastructure group. It plans to widen access to all Daybreak Red customers but has given no date.

What OpenAI reports

All of these figures are OpenAI's own, relayed by VentureBeat. None has been reproduced outside the company.

  • ExploitGym, which asks a model to exploit a specified weakness: 48.5%, against 32.1% for GPT-5.6 Cyber.
  • Web application test with no source code access: it found 40.3% of vulnerabilities, against 32% for GPT-6 Astra.
  • RemoteExploitBench, with known flaws in GitLab, PostgreSQL and Jenkins: 95% of tasks completed. Astra, running under the broader Daybreak Blue program, refused every task.
  • Scope control: in ExploitGym Honeypot it avoided unauthorized attacks in 99.93% of runs without safety prompts, against 94.49% for GPT-5.6 Cyber.
  • Real software: 81 previously unknown vulnerabilities in Firefox that Mozilla rated high severity, and in a Linux kernel setup, working exploits for 948 of 1,724 reproduced vulnerabilities.

VentureBeat's own caveats matter more than any single score. The 95% figure comes from a different evaluation than the August completion rate for GPT-5.6 Cyber, so matching headline numbers do not mean equal performance. The kernel exploit count is not a count of unknown or unpatched flaws. And there is no independently run, cross-vendor comparison.

How it compares on paper

Prices below are from the VentureBeat comparison, checked on 2026-11-05.

Model Input / output per M tokens Access
OpenAI GPT-6 Cyber $20 / $100 Selected Daybreak Red customers
OpenAI GPT-5.6 Cyber $12.50 / $75 All Daybreak Red customers
Anthropic Claude Mythos 5.1 $10 / $50 Vetted organizations, trusted-access programs
Google Gemini 4 Argon Cyber Permissive $4 / $20 Fairwind approval
Google Gemini 3.8 Flash Cyber $1.50 / $7.50 Approval list via Fairwind

GPT-6 Cyber's base rates are twice Mythos 5.1's and five times Argon Cyber Permissive's. Token price is not cost per finding, though. A model that completes an exploit in one pass can be cheaper than a cheap model that needs five attempts. Nobody outside the vendors has published that number for these models yet.

Who should care

If you do penetration testing, vulnerability research or run a security team at a large company, apply through Daybreak Red and run your own targets. A cyber model's value depends on your stack, so a vendor's benchmark tells you little about your code.

If you build a product that needs code review or bug triage, you don't need this model. Our ranking of the best LLMs for code review covers what you can call today, and our cost breakdown for agent workloads shows how to price a loop that retries. For general coding, see how Claude Opus 4.7 compared with GPT-5.5.

We'll retest if access opens, on code we control, and report cost per confirmed finding next to OpenAI's claims.

Related articles

LLM agent cost in production 2026: $0.40 for an 8-step loop
LLM agent costs in production 2026: three workload models with token math and USD totals per 1,000 runs, from a 3-call support bot to an 8-step coding loop.
Best LLM for code review in 2026: Haiku 4.5 beats GPT-4o
We tested four LLMs on six real buggy diffs: Claude Opus 4.7 swept the field, Haiku 4.5 beat GPT-4o 5-0, and GPT-4o finished with zero wins in 2026.