How Much To Run AI
Hosted API

GPT-6 Cyber: Daybreak Access, API Pricing and Task Costs

GPT-6 Cyber is OpenAI’s hosted model for investigating vulnerabilities, testing them in authorized environments and verifying fixes. Access requires Daybreak Red approval plus approval for this model. Standard short-context API rates are $20 per million input tokens and $100 per million output tokens; repeated tool calls and reasoning can make a complete investigation cost much more than its final written report.

Content reviewed:

Check GPT-6 Cyber access and specifications ↗
GPT-6 Cyber rates of $20 per million uncached input, $2 cached, $100 output and $25 or $50 cache writes, an example investigation turn at $3, and OpenAI’s Linux campaign at about $1,279 per patched vulnerability across 948 CVEs.
Original diagram of the rate card and published figures; the campaign average is not a per-audit price, and access requires Daybreak Red approval per project.

Use this model

See access details in the guide below.

Use cases

  • Investigate a suspected vulnerability in software you are authorized to assess.
  • Reproduce a finding in an isolated test environment and check a candidate fix.
  • Compare complete investigation costs with your existing security workflow.

Who can use GPT-6 Cyber?

OpenAI introduced GPT-6 Cyber on November 5, 2026, initially through the API and Codex for selected approved Daybreak Red customers. A general paid API account or ChatGPT subscription does not establish access. Confirm the exact model permission with your account team before changing an application’s model ID.

For the API, an organization owner must enable Daybreak Red on the calling project. With project-level controls, use a non-default project and a key belonging to that project. Add your application’s public egress IPs to the API IP allowlist. Daybreak Red organizations also require SSO and MFA through their identity provider. Projects using Zero Data Retention need validated Private Safety Processing and the matching retention policy.

The model ID is gpt-6-cyber. Its documented context window is 1,050,000 tokens, with up to 128,000 output tokens. It accepts text and image input and produces text. These limits describe capacity; choose a smaller, relevant evidence set when starting an investigation.

GPT-6 Cyber API prices and a task budget

The official Standard short-context prices checked November 9, 2026 are quoted per million tokens. Cache reads and explicit cache writes have separate rates. Use the pricing category that matches the request instead of applying a cache-read discount to all repeated text.

Token categoryUSD / million tokens
Uncached input$20
Cached input$2
Output$100
30-minute cache write$25
12-hour cache write$50

Count every investigation turn

A planning example with 100,000 uncached input tokens and 10,000 billed output tokens costs $3: $2 input plus $1 output. Ten requests of that size cost $30. This example has no cache writes; add any writes at the rates above. Hosted-tool calls, your own tool execution, sandbox hosting and applicable processing surcharges add to the token bill.

Reasoning tokens are billed as output even when they are absent from the final answer. Accumulate the usage from every request, including repeated context and tool results, then divide the full bill by accepted, independently verified findings or fixes.

OpenAI lists short-context rates for requests with up to 272K input tokens; it has not published rates above that limit for this model. For a larger request, check the current pricing page and your agreement before using this example as a budget. Codex plan allowances and API token billing are different purchasing paths.

Send a first approved API request

After the project and network setup is complete, set OPENAI_API_KEY locally or on your server to a key from the approved project. The following request follows OpenAI’s Daybreak guide. Replace the sample safety identifier with a stable pseudonymous identifier for the person using your application.

Inspect the returned model and usage, then read the generated test suggestion. This first request asks for text; executing a regression test requires your own test environment and tool loop. Keep actual reproduction work within the agreed target and permissions.

Request a regression-test suggestion
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 in an isolated test application."
  }'

What do the security results establish?

OpenAI reports 48.5% on ExploitGym versus 32.1% for GPT-5.6 Cyber, and 40.3% vulnerability recall on its offensive web-testing evaluation versus 32.0% for GPT-6 Astra. These measure different tasks and evaluation setups. They help identify a candidate for a security workflow; they do not determine the cost or success rate on your repository.

Begin with a bounded case containing a known bug and a patch with an independent regression test. Record accepted findings, failed attempts, human review, tool time and the complete token bill. Compare the same scope and verifier with your current model before increasing the investigation budget.

A published campaign cost, not a per-task quote

OpenAI reports $1,213,430 in API credits for its mid-October Linux run, which validated and generated patches for 948 distinct local privilege-escalation vulnerabilities: about $1,279 per patched vulnerability across the campaign. The run produced 1,169 working exploits, so dividing by exploit count gives a lower figure. Its earlier July campaign with GPT-5.6 Sol produced 166 local privilege-escalation exploits at $513,690. These used the same research harness but scanned the latest kernel at different dates. This is an OpenAI research comparison with Sol, not a matched external test against GPT-5.6 Cyber or a quoted price for an individual audit.

CodeRabbit: known issues caught versus findings reported

CodeRabbit’s November 5 security evaluation compares its ensemble with and without GPT-6 Cyber on the same 33 security cases, using high reasoning effort for investigation. The Cyber ensemble caught 22 selected known issues versus 19, and reported 361 high- or medium-severity findings versus 218. CodeRabbit says the additional findings still need assessment for validity; the 65.6% increase measures report volume, not verified extra vulnerabilities.

Some investigations needed recovery runs. The post does not publish complete token bills, scan times or a matched GPT-5.6 Cyber comparison. Use it to choose cases for your own validation, while budgeting review and recovery as well as model calls.

Before you use it

  • Access and rate limits apply to the approved organization and calling project; model approval does not transfer to another project.

How to get started

Confirm the approved project and outbound network, send the small regression-test request, then record full usage before increasing the investigation scope.

Sources

Sources (6)

Calculators