GPT-6 Cyber costs twice as much as Claude Mythos 5.1, and only OpenAI-vetted teams get in

Twenty dollars per million input tokens and a hundred per million output: GPT-6 Cyber, the security model OpenAI launched on November 5, 2026, costs twice as much as Claude Mythos 5.1 and five times as much as the Cyber build of Gemini 4 Argon. In return it promises to find and exploit more vulnerabilities than any earlier OpenAI model, and only organisations OpenAI approves one by one can use it.

01What it is and who gets it

GPT-6 Cyber is a GPT-6 Astra variant with extra training for investigating vulnerabilities, reproducing them in an authorised environment and verifying the fix. It is in the API as gpt-6-cyber, through Responses and Chat Completions, with a 1,050,000-token window (922,000 input tokens at most) and up to 128,000 output tokens. Batch and fine-tuning are not supported.

There are two gates. Your organisation needs Daybreak Red, OpenAI's programme for authorised offensive work, plus a separate approval for this model. According to VentureBeat, only some Daybreak Red customers hand-picked by OpenAI get it at first; the programme had more than 2,000 enrolled entities as of October.

02How much better, and at what price

OpenAI's numbers, as reported by VentureBeat: on ExploitGym (exploiting a specified weakness) it goes from 32.1% for GPT-5.6 Cyber to 48.5%; on a web-application test without source code it finds 40.3% of vulnerabilities, against 32% for GPT-6 Astra. On the scope test, which checks whether the model wanders off its assigned target to attack nearby services, it stayed in scope in 99.93% of runs without safety prompts (94.49% for GPT-5.6 Cyber).

Prices per million tokens. As a rough guide, a review that reads 2 million tokens of code and writes a 200,000-token report costs about $60 on GPT-6 Cyber, $40 on GPT-5.6 Cyber and $30 on Mythos 5.1, before caching. Cache writes are billed separately: $25 per million for 30 minutes, $50 for 12 hours.

What is missing is a head-to-head with Mythos 5.1 or Gemini on the same benchmark: OpenAI only compares against its own models. Being twice the price of Mythos tells you nothing about which one finds more bugs per dollar.

03Turning it on

If your organisation already has approval, the docs ask for three things before the first call:

• An owner enables Daybreak Red on the project (not the default one), and you use a key from that project.
• An IP allowlist with the public IPs your application calls from.
• If the project uses Zero Data Retention, Private Safety Processing must be configured: OpenAI reviews encrypted records that you store for 30 days, and it applies to all traffic in the project.

The minimal call, from the Daybreak guide:

curl https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6-cyber",
    "access_programs": {"cyber": "daybreak_red"},
    "safety_identifier": "user_a17f",
    "input": "Suggest a regression test for a path traversal fix."
  }'

04The gotchas

The easiest one to trip on: access_programs.cyber follows the model, not your approval level. gpt-6-sol, gpt-6.1-sol and gpt-6-astra need daybreak_blue even if you hold Red, and return invalid_access_program with daybreak_red; gpt-6-cyber and gpt-5.6-cyber, on the other hand, take daybreak_red.

The other one has a date: gpt-5.5-cyber-preview retires on February 1, 2027, and requests to it will fail from then on. OpenAI suggests moving to gpt-5.6-cyber, or to gpt-6-cyber if you already have approval. If a security-review pipeline of yours runs on the preview model, that is the first thing I would change: the model id is one line of config, but the replacement's price and behaviour are not the same, and you want to measure them on your own cases before February.

05Sources

• GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents (OpenAI, November 5, 2026)
• GPT-6-Cyber Model (OpenAI API docs)
• Use Daybreak in the Responses API (OpenAI API docs)
• OpenAI launches GPT-6 Cyber with stronger exploit research and higher API price than Anthropic and Google (VentureBeat, November 5, 2026)