OpenAI added gpt-6-cyber to its API changelog on November 5, 2026. The model page prices it at $20 per million input tokens and $100 per million output tokens, and says an organization needs Daybreak Red access and approval to use it. The Daybreak guide is more exact: Red access alone does not cover this model, which needs additional model approval.
Everything below is read from OpenAI’s developer documentation. S5 Labs has not used or tested the model, and the developer pages we read give no capability or benchmark figures, so there are none here. Our August Daybreak explainer covers the Blue and Red tiers. This piece adds the prices, setup gates and changelog date it lacked.
The model page lists two endpoints and no Batch support
The only model ID listed is gpt-6-cyber. The page gives a 1,050,000-token context window, 128,000 max output tokens and reasoning-token support, with text and image input and text output. Two endpoints are supported, Chat Completions and Responses. Batch, Realtime, Assistants and fine-tuning are among those shown as unsupported. Listed features are streaming, function calling, structured outputs, web search and prompt caching, and the Responses API tool list includes hosted shell, code interpreter, computer use and MCP. The page shows no knowledge cutoff and no cloud-provider model IDs. GPT-6-Cyber model page
The input rate doubles past 272K input tokens
Standard prices per 1M tokens:
| Rate | Requests up to 272K input tokens | Requests over 272K input tokens |
|---|---|---|
| Input | $20 | $40 |
| Cached input | $2 | $4 |
| Output | $100 | $150 |
| Cache write, 30 minutes | $25 | not listed |
| Cache write, 12 hours | $50 | not listed |
GPT-6-Cyber model page · API pricing
This is the same step we described for GPT-6.1 Terra: past 272K input tokens, input and cached input double and output goes to 1.5x. Our Terra article lists a half-price rate for Batch and Flex on that model. The GPT-6-Cyber page marks the Batch endpoint as not supported, and the pricing page’s Batch and Flex tables do not list the model, so we found no discounted rate for it.
The next table is S5 Labs arithmetic from those rates: Standard pricing, uncached input, no cache writes, no tool fees. It assumes the higher rates cover the whole request once input passes 272K, which the gpt-5.6-cyber page states for its own step and the gpt-6-cyber page only implies. Confirm against a real invoice.
| Request | Input | Output | Total |
|---|---|---|---|
| 100,000 in, 10,000 out | $2.00 | $1.00 | $3.00 |
| 272,000 in, 20,000 out | $5.44 | $2.00 | $7.44 |
| 273,000 in, 20,000 out | $10.92 | $3.00 | $13.92 |
| 800,000 in, 50,000 out | $32.00 | $7.50 | $39.50 |
On those assumptions, one thousand extra input tokens take the second row from $7.44 to $13.92, an 87% increase. An agent loop that keeps appending tool output will cross 272K without anyone choosing to, so cap input per request or budget the whole job at the long-context rate.
The first row is the sample request from our Terra article, where it cost $0.15 on Terra and $1.50 on Astra. On gpt-5.6-cyber, at $12.50 input and $75 output, it costs $2.00. GPT-6-Cyber is 60% more expensive on input and a third more on output than the model already available to every Daybreak Red customer. That model’s context window is 400,000 tokens, against 1,050,000 for GPT-6-Cyber. GPT-5.6-Cyber model page
A cached read costs a tenth of the input rate. If a cache write is billed in place of ordinary input, which the page does not spell out, a 30-minute write pays for itself on the first reuse and a 12-hour write on the second.
The page’s Standard rate-limit table runs from 3 requests and 10,000 tokens per minute on Free through 5,000 and 1,000,000 on Build to 15,000 and 40,000,000 on Grow. It says the limits apply to approved organizations and may differ if yours has custom limits. Build’s token limit is smaller than the context window, so ask about custom limits before planning long-context work there. GPT-6-Cyber model page
Five gates before the first request
| Gate | What the docs require |
|---|---|
| Approval | Daybreak Red access plus approval for GPT-6-Cyber. Confirm both with your account team. |
| SSO and MFA | Every Daybreak Red organization must require SSO for the API organization that holds Cyber access and enforce MFA through its identity provider. With a separate Cyber organization, verify its domain and test SSO there. |
| Project | An organization owner enables Daybreak Red for the project that will send requests. With project-level controls, use a non-default project and a key from it. |
| IP allowlist | To use GPT-6-Cyber through the API, allowlist the public IPs your application sends requests from. If you are approved to embed the model in a product, end users without direct model access do not need to allowlist their own IPs. |
| Zero Data Retention | If the project uses ZDR, configure Private Safety Processing (PSP). Storage must show “Validated” and the retention policy must read “Zero Data Retention with Private Safety Processing”. |
The approval row is where this adds to our August piece, which described Red as the gate to GPT-5.6-Cyber. That still holds for gpt-5.6-cyber, which its model page says is available to all Daybreak Red customers. For GPT-6-Cyber, the docs give no procedure or lead time for the extra approval beyond asking the account team. The cybersecurity-checks guide adds that applying, verifying an identity or receiving Daybreak Blue access does not grant specialist-model access.
The same guide says Trusted Access for Cyber, which it describes as a reviewed access program and not the name of a model, does not automatically grant Zero Data Retention. Per the Private Safety Processing guide, PSP is enabled per project and then applies to all API traffic in that project, so keep ZDR workloads that should run without it in a separate project.
According to that guide, setup needs a ZDR-approved organization, storage in AWS S3, Azure Blob, Google Cloud Storage or OCI Object Storage, and an organization Admin API key for Management API calls. Interactions selected by a safety-classifier referral or an approved sampling policy are encrypted and written to your regional storage with a 30-day TTL. “Validated” records one successful check, not ongoing storage health. Disconnecting the project’s last storage connection resets its retention policy to the organization default, and models that require ZDR with PSP may then become unavailable.
Anthropic and Google gate their cyber models under their own rules, which we did not recheck here; see our pieces on Anthropic’s Cyber Verification Program tiers and Google’s Gemini 4 Carbon.
The request value follows the model
The guide’s GPT-6-Cyber example is a POST /v1/responses call with this body:
{
"model": "gpt-6-cyber",
"access_programs": {"cyber": "daybreak_red"},
"safety_identifier": "user_a17f",
"input": "Suggest a regression test for a path traversal fix."
}
access_programs.cyber selects behavior within access you already have and does not grant any. The value to send depends on the model, whatever your organization’s approval level:
| Model | Approval required | Value to send |
|---|---|---|
| GPT-6 Sol | Blue or Red | daybreak_blue |
| GPT-6.1 Sol, GPT-6 Astra | Red | daybreak_blue |
| GPT-5.6-Cyber, GPT-5.5-Cyber (existing access only) | Red | daybreak_red |
| GPT-6-Cyber | Red + additional model approval | daybreak_red |
A Red-approved organization that sends daybreak_red to gpt-6-sol gets 400 invalid_access_program, and the error names the value to use. For 400 unsupported_access_program, the fix is a Daybreak-capable model or the value standard. 403 access_program_not_enabled means the key’s project lacks the program, or organization approval or project access is missing.
The guide’s omission defaults do not name gpt-6-cyber. One covers Red models “such as” gpt-5.6-cyber, for which the API selects daybreak_red. Another says that starting with gpt-6.1-terra, omitting the parameter results in standard treatment. So set daybreak_red explicitly, as the official example does. The guide is also written for the Responses API. The model page lists Chat Completions as supported, but nothing we read says whether access_programs works there, so start on Responses.
Per-user identifiers limit who gets cut off
OpenAI’s cybersecurity-checks guide says the company classifies GPT-6-Cyber as Critical for cybersecurity under its Preparedness Framework, and that additional automated safeguards apply to API models rated High or Critical. When suspicious activity passes OpenAI’s thresholds, requests return an error with code cyber_policy and access may be temporarily revoked. Without a per-user safety_identifier, that can cover the whole organization. With a unique identifier per end user, OpenAI says access may be revoked for the affected user instead, after human review and warnings. Cybersecurity checks
So replace the sample user_a17f with a stable, pseudonymous ID for whoever made the request, as the Daybreak guide instructs. For ZDR organizations, request-level mitigations also apply, and on streaming requests the error may arrive in the middle of other events. The appeals section says to contact support if you believe access was limited incorrectly and need it restored before a 7-day period ends. The page does not define that period.
gpt-5.5-cyber-preview leaves the API on February 1, 2027
The deprecations page, under a heading dated November 5, 2026, says gpt-5.5-cyber-preview will be removed from the API on February 1, 2027. It names no specific replacement, only the most capable cyber model available to each customer. The Daybreak guide says requests fail after retirement, so change the model to gpt-5.6-cyber before then, or to gpt-6-cyber if you have approval, and confirm the project can reach the replacement first.
Without the extra approval, the replacement is gpt-5.6-cyber. It supports the Responses endpoint only, so keep any fallback from GPT-6-Cyber on Responses too.
What to test once you have access
- Send the guide’s example from the approved project and allowlisted IP, and log the
access_programs.cyberfield in the response, which records the selected program when available. - Log input tokens per request for a week of real work and count how many pass 272K before you set a budget.
- Open a streaming request and make sure your client handles a
cyber_policyerror mid-stream. - Try
access_programson Chat Completions in a test project before production code depends on it. - Run your own tasks on
gpt-5.6-cyberandgpt-6-cyberside by side. The docs give no measurement showing the newer model is worth the higher price.
