OpenAI Daybreak Blue and Red Split the API in Two: What Every Gateway Operator Must Audit Now

OpenAI shipped two new model IDs on August 7 under its Daybreak cybersecurity program: daybreak-blue-latest and gpt-5.6-cyber. Coverage has mostly focused on what the models do — GPT-5.6 Sol for defensive work, purpose-trained Cyber for exploit validation and red teaming. What it has not focused on is what Daybreak does to your gateway layer.

The short version: both models are Responses API only. v1/chat/completions is explicitly unsupported. Batch is unsupported. Fine-tuning is unsupported. If your team or your customers route security tooling through a standard OpenAI-compatible proxy — Cursor, Claude Code with a custom gateway endpoint, any SDK calling /v1/chat/completions — that traffic fails immediately for these model IDs, even with a valid API key.

What Daybreak actually ships

Daybreak Blue (daybreak-blue-latest) is an alias model that defaults to GPT-5.6 Sol. It is designed for broad defensive security workflows: vulnerability discovery, secure code review, detection engineering, incident response, malware analysis, and patch validation. Context window is 1,050,000 tokens, maximum input 922,000 tokens, maximum output 128,000 tokens.

Daybreak Red (gpt-5.6-cyber) is a purpose-trained cybersecurity model with a narrower context window: 400,000 tokens context, 272,000 maximum input, 128,000 maximum output. Pricing is $12.50 per million input tokens and $75 per million output tokens — roughly 3× the input and 10× the output price of GPT-5.6 Sol at standard tier. Above 272K input tokens, pricing doubles on input and increases by 50% on output.

Both models support streaming, structured outputs, function calling, file search, image input, web search, prompt caching, code interpreter, hosted shell, apply_patch, skills, computer use, and mc. Both are Responses API only.

The provisioning gate that most operators will miss

Access to Daybreak is not unlocked by having an existing OpenAI API key. Daybreak requires separate approval and provisioning through the Daybreak Access form. An organization on a standard Enterprise or API tier receives a 401 or 404 when calling these model IDs without approved access, regardless of billing status.

This creates a provisioning architecture problem that is worth naming explicitly. A gateway operator running a shared endpoint across multiple customer organizations cannot route Daybreak traffic on behalf of those organizations unless each organization has separately applied to and been approved for Daybreak access. There is no delegation path from a single approved Daybreak account to downstream customers. The model requires customer-level approval, not gateway-level approval.

That is structurally different from how most routing works. With GPT-5.6 Sol, Azure AI Foundry, or Anthropic's standard Claude API tiers, a gateway operator authenticates with a single upstream key and routes requests on behalf of its customers. Daybreak breaks that topology.

Endpoint incompatibility is not optional to work around

The Responses API and Chat Completions API are not the same endpoint. Chat Completions (v1/chat/completions) is the API that OpenAI-compatible SDKs, Cursor, Claude Code, and nearly every third-party integration uses by default. The Responses API (v1/responses) is newer and has a different request shape — it uses input instead of messages, text.format instead of response_format, and has its own tool call structure.

An operator who points an OpenAI SDK at a gateway endpoint and requests gpt-5.6-cyber will get an endpoint error, not a model error. The SDK is likely calling /v1/chat/completions. The model does not respond there. Adapting requires either migrating the SDK to the Responses API natively, or the gateway implementing a translation layer — which in turn requires the gateway to specifically handle Daybreak model IDs and route them differently.

The practical decision tree for a gateway operator today:

1. Are any of your customers running authorized security workflows that would benefit from Daybreak?
2. If yes, do those customers have Daybreak approval, or are they expected to apply?
3. If yes, are their SDKs calling the Responses API, or Chat Completions?
4. If Chat Completions, does your gateway translate to Responses API for specific model IDs?

Most gateway operators will answer no to step 1 and stop there. But for those building security tooling platforms or serving security teams, the answers to steps 2 through 4 determine whether Daybreak is accessible at all.

How this compares to Anthropic's gated-model architecture

Anthropic has its own gated-tier precedent: Fable 5 (released on a separate track from Claude Sonnet and Opus), and earlier the Astra "Critical Cyber" designation that made headlines in August. Anthropic's gated models, however, use the same Messages API endpoint. A Fable model ID works over the same /v1/messages path that all Claude traffic uses. Gating happens at authorization time, not at endpoint routing time.

OpenAI's Daybreak does something more architecturally disruptive by making the endpoint itself gated. Daybreak Blue and Red only answer at v1/responses. This means that Daybreak traffic is routed differently not just in policy, but in HTTP path and request schema.

From a routing policy perspective, that has a cleaner upside: a gateway can route Daybreak model IDs to the Responses API endpoint specifically, and route everything else to Chat Completions, without needing to inspect authentication at the routing layer. The model ID itself disambiguates the endpoint. The downside is that the SDK layer below the gateway needs to support both endpoints, or the gateway needs to translate request shape.

What to watch over the next 90 days

OpenAI's Daybreak Cyber Partner Program is expanding. The current partner list includes enterprise security vendors, and the Codex Security integration deepens the connection between Daybreak models and agentic security workflows.

If you are building on top of OpenAI's Responses API today — for any use case, not only security — Daybreak is worth tracking as a pattern. The Responses API has been gaining capabilities (retained reasoning, stored outputs, computer use tools) that Chat Completions has not received. Daybreak being Responses-API-exclusive is consistent with OpenAI positioning the Responses API as the forward-looking endpoint and Chat Completions as legacy.

For operators: if a security team in your customer base asks about Daybreak model IDs in your gateway, the first question to resolve is not pricing — it is whether they have Daybreak provisioning, and whether your gateway handles v1/responses translation.

What TheRouter users should watch

TheRouter routes OpenAI-compatible traffic across configured providers. For standard OpenAI model IDs, routing works over /v1/chat/completions. Daybreak model IDs are not part of standard OpenAI-compatible routing and require the Responses API endpoint plus separate provisioning approval.

If you are evaluating Daybreak for a security use case, verify Daybreak provisioning directly with OpenAI before building your gateway configuration. The OpenAI Responses API docs describe the endpoint shape; the Daybreak-specific endpoint restrictions are documented at developers.openai.com/api/docs/models/daybreak-blue-latest and /gpt-5.6-cyber.