GPT-6 Cyber: Who Can Use It and How Daybreak Access Works

GPT-6 Cyber arrived on November 5, 2026, the date on OpenAI's system card. Access runs through an organization approved for Daybreak Red, with a separate approval for the model and a project with the program enabled. Daybreak Blue approval on its own does not unlock GPT-6 Cyber.

Those are three separate permissions: the program, the model, and the project and key. The access_programs.cyber request value selects behavior inside that scope; it does not grant access.

Can you use GPT-6 Cyber today?

The model card says your organization needs Daybreak Red access and approval for GPT-6 Cyber, and your account team can confirm both (checked November 6, 2026). OpenAI reviews identity, security practices and legal attestations, and completing verification does not guarantee approval.

Ordinary organizational approval covers authorized internal work, not external users or third-party customers. OpenAI separately approves embedded products, and end users of an approved product can use the model without direct model access.

Daybreak Red vs Blue: approval and request values

Red and Blue are access programs, not models and not plan tiers. OpenAI's trusted-access page describes Blue as access to flagship models with reduced refusals for authorized defensive work, and Red as separately approved access to specialist cyber models for more advanced security research.

Model approval and the program value are separate. This table comes from the Daybreak guide (checked November 6, 2026):

See the GPT-6 Astra and GPT-6.1 Sol model pages for model details.

The field accepts three exact strings: standard, daybreak_blue and daybreak_red. Read the right-hand column as the value for that model, not your organization's highest approval.

GPT-6 Astra and GPT-6.1 Sol are the counterintuitive pair. Red approval is what turns on reduced refusals, while the request value stays daybreak_blue; both models reject daybreak_red, and Blue approval alone does not authorize those reduced refusals on either one. Both models also run under standard safeguards without Daybreak.

If you omit the field, mainline models such as gpt-6-sol get Blue treatment when the organization and project are entitled, and standard safeguards otherwise. Red models such as gpt-5.6-cyber select daybreak_red and fail when access is missing. From gpt-6.1-terra onward, omitting it gives standard treatment.

What changed with GPT-6 Cyber?

GPT-6 Cyber is a cyber-specialized variant of GPT-6 Astra, built for authorized security work such as vulnerability discovery, exploit development and web penetration testing. The system card gives it the same internal Preparedness designations as Astra, including Critical in cybersecurity, and OpenAI says it is better aligned than GPT-5.6 Cyber.

OpenAI's own evaluation, reported in its system card, ran at the highest tested reasoning budget. On SEC-Bench Pro, Cyber scores 87.02% pass@1 for vulnerability discovery in large JavaScript engines, against 85.38% for GPT-6 Astra (system card, checked November 6, 2026).

On the model card, gpt-6-cyber takes text and image input and returns text. Audio and video are listed as unsupported, and so is fine-tuning. Streaming, function calling, structured outputs and the hosted tools in the Responses API are listed as supported.

Set up an approved project and send a request

Prepare the approved organization and project as follows, based on the Daybreak guide (checked November 6, 2026):

1. Confirm with your account team that the organization has Daybreak Red and approval for GPT-6 Cyber.
2. Ask an organization owner to enable Daybreak Red on the project that will send requests. With project-level controls, use a non-default project and a key from it.
3. Configure an IP allowlist for the public addresses your application sends from.
4. Require SSO for the API organization that holds Cyber access and enforce MFA through your identity provider. If you use a separate Cyber organization, verify its domain and test SSO there.
5. If the project uses Zero Data Retention, configure Private Safety Processing and confirm the retention policy.
6. Send the request with model gpt-6-cyber and access program daybreak_red.

This is the request shape from OpenAI's setup example, with its patch-validation prompt. Replace the sample safety identifier with a stable pseudonymous ID for the end user. The example follows OpenAI documentation and passed shell/JSON syntax checks; it has not been executed against the API.

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."
  }'

The prompt asks for a regression test, so the work stays on validating a fix in an authorized test environment.

When a program applies, the response records it in the access_programs.cyber field. When standard safeguards ran, access_programs is null. With a latest alias, read the resolved model field, because alias resolution depends on your approved access.

If the request fails, match the status and code:

• 400 invalid_access_program: the model needs a different program value. Change the access program value to the one named in the error and retry.
• 400 unsupported_access_program: that model does not take the program. Switch to a Daybreak-eligible model, or send standard.
• 403 access_program_not_enabled: the error can mean the organization lacks the required approval, or the key's project lacks the program. Check the project and key first; if project access is missing, ask an organization administrator to enable it. If organization approval is missing, request the level named in the error.

Cyber safeguards can also return the error code cyber_policy and temporarily limit access; the cybersecurity checks guide describes a 7-day window to request restoration.

Official API pricing and limits

GPT-6 Cyber costs $20 per million input tokens and $100 per million output tokens at OpenAI's standard short-context rates. Cached input is $2 per million, and cache writes are $25 per million for 30 minutes or $50 per million for 12 hours (model card, checked November 6, 2026).

At those rates, one request with 50,000 input tokens and 10,000 output tokens costs $2.00: $1.00 for input and $1.00 for output. The same request on GPT-6 Astra is $1.00, and on GPT-6.1 Sol it is $0.20.

All figures are OpenAI list prices in US dollars per million tokens, not OmniaKey prices, from the model card and the pricing page.

The model card lists a 1,050,000-token context window and a maximum of 128,000 output tokens. OpenAI publishes standard short-context prices for this model.

What to use for ordinary coding work

GPT-6 Cyber is designed and gated for specialized authorized security work. For ordinary product development, compare general models. OmniaKey's catalog lists GPT-6.1 Sol and GPT-6 Astra, but Daybreak approval is not part of an OmniaKey key.

The GPT-6 Astra review covers the general model in detail, and the Astra setup guide covers the standard API call.

For family selection, the GPT-6 Sol vs Astra comparison covers the trade-offs.

Frequently asked questions

Does Daybreak Blue approval unlock GPT-6 Cyber?

No. Blue and Red are separate reviews, and OpenAI's trusted-access page says applying for or receiving Blue does not grant Red. GPT-6 Cyber also needs its own model approval on top of Red.

Which program value do I send with GPT-6 Astra or GPT-6.1 Sol?

Send daybreak_blue. Both models require Red approval for reduced refusals, and both reject daybreak_red.

What does a 400 invalid_access_program error mean?

The model needs a different program value. Change the access program value to the one named in the error and retry.

How much does one GPT-6 Cyber request cost?

At standard short-context rates, 50,000 input tokens and 10,000 output tokens cost $2.00. Cached input is much cheaper at $2 per million tokens.

Sources checked

• OpenAI: GPT-6 Cyber model card
• OpenAI: Use Daybreak in the Responses API
• OpenAI: GPT-6 Cyber system card
• OpenAI: API pricing
• OpenAI: Cybersecurity checks
• OpenAI: Trusted access for cyber

Snapshot: November 6, 2026, America/New_York. OpenAI's Daybreak guide and the GPT-6 Cyber model card are the two pages to recheck when approvals or prices change.