GPT-6-Cyber API Guide: Setup, Daybreak Red, and Migration

Learn how to configure OpenAI's GPT-6-Cyber model, handle Daybreak Red access programs, configure network security, and migrate before older previews retire.

By ailog · Published Nov 8, 2026 · 8 min read · How we work
In short
  • OpenAI released the gpt-6-cyber model on 2026-11-05 for defensive vulnerability reproduction, fix verification, and research in sanctioned environments.
  • Access mandates Daybreak Red qualification along with supplementary model validation, IP allowlisting, and mandatory identity provider SSO with MFA.
  • The architecture delivers a 1,050,000 context window, 922,000 maximum input tokens, and 128,000 output tokens across the Chat Completions and Responses endpoints.
  • Standard pricing charges $20 per 1M input tokens, $2 per 1M cached input tokens, and $100 per 1M output tokens, with cache write tiers at $25 and $50.
  • Teams operating gpt-5.5-cyber-preview must transition to gpt-5.6-cyber or gpt-6-cyber ahead of its official retirement on February 1, 2027.

What Is GPT-6-Cyber and What Changed

On 2026-11-05, the OpenAI API changelog announced the formal release of gpt-6-cyber. This specialized model is purpose-built to assist technical security teams in investigating vulnerabilities, reproducing defects inside authorized test beds, and validating functional patches before production release. Unlike general-purpose reasoning models, gpt-6-cyber operates under the strict governance of the OpenAI Daybreak access framework, enabling authorized organizations to work through complex software security problems with fewer generic refusals.

The release accompanies a structured migration schedule for older defensive models. Teams still maintaining workloads on gpt-5.5-cyber-preview must transition to either gpt-5.6-cyber or gpt-6-cyber because gpt-5.5-cyber-preview will retire on February 1, 2027. Requests sent to that retiring snapshot after that cutoff date will fail. For engineering teams evaluating model options across providers, reviewing comparative infrastructure costs in the LLM API cost calculator or examining model deprecations across providers provides broader context on operational lifecycles.

Organization Prerequisites and Onboarding Setup

Gaining access to gpt-6-cyber involves strict governance criteria that extend beyond conventional API tiers. An organization must hold Daybreak Red approval and receive additional model-specific authorization confirmed by their OpenAI account team. Simply having standard Daybreak Blue credentials is not sufficient to call this model.

  • Single sign-on enforcement: Every Daybreak Red organization must enforce enterprise SSO and mandate MFA through their chosen identity provider for the API organization holding cyber privileges. If using an isolated Cyber organization, you must verify the organization domain and validate SSO operations there.
  • Dedicated project keys: An organization owner must enable Daybreak Red for the designated project issuing the requests. Operators must generate secret keys within that specific, non-default project rather than using global workspace credentials.
  • Public IP allowlisting: Applications routing traffic directly to gpt-6-cyber must originate from explicit public network ranges defined in the project IP allowlist. However, downstream end users who interact with an embedded security product do not need to allowlist individual personal IP addresses.
  • Zero Data Retention rules: If your organization operates under Zero Data Retention (ZDR), you must configure Private Safety Processing (PSP). Project storage must indicate a status of Validated, and the active retention tier must show Zero Data Retention with Private Safety Processing. This setting governs every API interaction routed through the project.

Teams deploying strict infrastructure boundaries can review network strategies like Anthropic IP allowlisting setup or general enterprise controls like API data retention and training policies to align enterprise compliance frameworks.

Configuring the Responses API and Access Program Flags

Calling gpt-6-cyber requires passing specific flags to signal approved program participation. In the Responses API, callers use the access_programs.cyber field. For gpt-6-cyber, the valid program identifier is daybreak_red. It is critical to note that the API value must match the requirements of the chosen model rather than the broad approval level of the organization.

Model Approval Required API Value (access_programs.cyber)
GPT-6 Sol Blue or Red daybreak_blue
GPT-6.1 Sol Red daybreak_blue
GPT-6 Astra Red daybreak_blue
GPT-5.5-Cyber (existing access only) Red daybreak_red
GPT-5.6-Cyber Red daybreak_red
GPT-6-Cyber Red + additional model approval daybreak_red

For example, while gpt-6.1-sol and gpt-6-astra require an organization to hold Daybreak Red clearance to obtain reduced refusals, the request payload must declare daybreak_blue; passing daybreak_red to either of those two models produces an error. Conversely, requests sent to gpt-6-cyber must specify daybreak_red.

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."
  }'

Both access_programs and its child field cyber are optional structures in standard requests, but neither parameter accepts a null literal value. Supplying an empty access_programs dictionary leaves request selection unspecified. When constructing calls, provide a stable, pseudonymous string in safety_identifier to identify the actor initiating the operational audit.

Model Specs, Context Windows, and Supported Endpoints

The architectural envelope of gpt-6-cyber accommodates large-scale codebase investigations, binary analysis, and patch diffing. The model supports text and image input modalities and produces text output, complete with reasoning token support. It offers an expansive 1,050,000 token context window, accepting up to 922,000 maximum input tokens and generating up to 128,000 maximum output tokens per single completion.

  • Active snapshot: The default snapshot and locked identifier is gpt-6-cyber.
  • Supported endpoints: Requests are supported solely through Chat Completions (v1/chat/completions) and Responses (v1/responses).
  • Unsupported endpoints: Live sessions, Realtime audio, Realtime translation, Realtime transcription, Assistants, Batch, Fine-tuning, Embeddings, Image generation, Image edit, Videos, Audio speech, Audio transcriptions, Audio translations, Moderation, and legacy Completions are not supported.
  • Integrated toolset: Supported tools under Responses include function_calling, web_search, file_search, tool_search, image_generation, code_interpreter, hosted_shell, apply_patch, skills, computer_use, and mcp.
  • Supported platform capabilities: Native capabilities encompass streaming, structured_outputs, function_calling, web_search, and prompt_caching.

Because batching through v1/batch is not supported for gpt-6-cyber, engineering teams cannot offload deep inspection workloads onto asynchronous queues. High-throughput security scanning must instead be managed against live concurrency windows using streaming or standard response payloads. For details on how other platforms structure tools, see function and tool calling across providers or prompt caching mechanics.

Token Pricing and Rate Limits Across Usage Tiers

Understanding token expense is essential when feeding large codebases into an active audit pipeline. Pricing for gpt-6-cyber falls under standard short-context rates, with distinct costs assigned to prompt inputs, memory reuse, and generated output tokens.

Metric Price Unit
Input $20 1M tokens
Cached input $2 1M tokens
Cache writes (30 minutes) $25 1M tokens
Cache writes (12 hours) $50 1M tokens
Output $100 1M tokens

At $20 per 1M input tokens and $100 per 1M output tokens, this model commands a substantial premium over baseline general models such as gpt-6.1-sol ($2 input, $10 output per 1M tokens) or gpt-6-astra ($10 input, $50 output per 1M tokens). When reusing large security frameworks or dependency trees across repeated turns, prompt caching dramatically cuts ongoing ingestion expenses to $2 per 1M tokens, though teams must account for initial cache write fees: $25 per 1M tokens for short-context 30 minutes retention, or $50 per 1M tokens for 12 hours retention. Tool invocations that call external mechanisms such as search or computer use incur extra tool-specific fees.

Tier RPM TPM IPM
Free 3 10,000 50,000
Build 5,000 1,000,000 50,000
Launch 10,000 4,000,000 50,000
Grow 15,000 40,000,000 250,000

Throughput ceilings escalate according to usage tiers. While Free tier limits are strictly restricted to 3 RPM and 10,000 TPM, production organizations reaching Grow status can scale workloads up to 15,000 RPM, 40,000,000 TPM, and 250,000 IPM. To explore how token quotas differ across platforms, review rate limits and usage tiers.

Migration Path from GPT-5.5-Cyber and Error Handling

Organizations with grandfathered access to gpt-5.5-cyber-preview face an impending hard deadline: the snapshot retires on February 1, 2027. Teams should migrate their codebases well ahead of that date. If your organization lacks immediate approval for gpt-6-cyber, adjust the request model parameter to gpt-5.6-cyber. If your organization holds Daybreak Red credentials and secures the supplemental approval, transition straight to gpt-6-cyber.

  • 400 invalid_access_program: Emitted when the model and access flag do not align. For instance, sending daybreak_red to gpt-6-sol or gpt-6.1-sol triggers this fault. Update the access_programs.cyber value to the string indicated in the error message and retry.
  • 400 unsupported_access_program: Returned when requesting a Daybreak program on a model that does not offer cybersecurity flags. To resolve, point the request to a supported model or switch the program value to standard.
  • 403 access_program_not_enabled: Occurs when the API key belongs to a project without the requisite entitlement enabled, or if organization-level onboarding remains incomplete. Check that the API key was generated inside the approved non-default project and verify approval tiers with your workspace administrator.

Remember that omitting access_programs.cyber entirely can produce unexpected baseline behaviors depending on the target architecture. Starting with gpt-6.1-terra, omitted cyber flags fall back directly to standard handling. On mainline models like gpt-6-sol, omitting the field applies Daybreak Blue safeguards only if the project and organization already hold credentials. On dedicated Red engines like gpt-5.6-cyber, the gateway automatically selects daybreak_red and fails if authorization is missing. When using gpt-6-cyber, explicitly declaring {"access_programs": {"cyber": "daybreak_red"}} prevents ambiguous state resolution.

Implementation Checklist for GPT-6-Cyber

Before pushing automated security scripts or interactive patch testing tools into your testing pipelines, review the operational steps below to avoid authorization drops or runtime request failures.

  • Confirm Daybreak Red approval and obtain dedicated model authorization for gpt-6-cyber from your OpenAI account representative.
  • Enforce enterprise SSO and identity provider MFA across the organizational tenant holding cyber permissions.
  • Create a dedicated, non-default project and ensure an organization owner activates Daybreak Red for that specific workspace.
  • Register your outbound production gateway IPs in the project IP allowlist before issuing curl commands or SDK calls.
  • If Zero Data Retention is active, configure Private Safety Processing (PSP) and verify that storage displays Validated.
  • Update integration payloads to target v1/chat/completions or v1/responses, ensuring you pass daybreak_red in access_programs.cyber.
  • Incorporate a stable, pseudonymous string into the safety_identifier parameter on every outbound execution.
  • Audit legacy repositories targeting gpt-5.5-cyber-preview and migrate endpoints to gpt-5.6-cyber or gpt-6-cyber before February 1, 2027.
Sources
  1. OpenAI API changelog — 2026-11-05
  2. GPT-6-Cyber
  3. Use Daybreak in the Responses API

Related