OpenAI has added a new API surface for applications that need fast, structured judgments rather than open-ended generation. The company’s API changelog lists an Oct. 6 release for v1/decisions, a public beta endpoint that currently runs on GPT-6 Luna.
The practical distinction is important. OpenAI describes the Decisions API as a way to turn text and images into typed answers, and says it can be up to 10 times faster than using the Responses API for the same kind of classification-style work. Vercel added support for the new endpoint to its AI Gateway the next day, using the model identifier openai/gpt-6-luna-decisions.
That makes the launch more than a model-catalog update. It introduces a request and response pattern that sits alongside chat, generation, embeddings and tool-running. For teams operating a multi-model API gateway for teams, the question is no longer only which model to call. It is also which API shape best fits the task.
What changed
The Decisions API is aimed at workloads where the application needs a constrained result: a label, category, routing decision, policy outcome or other typed answer. Instead of asking a general-purpose model to produce prose and then parsing the result, developers can use an endpoint designed around structured decision output.
OpenAI says GPT-6 Luna is currently the only available model for the beta. That limitation matters because it means developers are not choosing among a broad family of decision models yet. They are testing a new OpenAI endpoint and a specific model binding at the same time.
The timing also shows how quickly gateway providers are expected to absorb new API surfaces. Vercel’s Oct. 7 support means developers using its gateway can call the Decisions API without integrating directly against OpenAI’s endpoint. Other OpenAI-compatible platforms will face the same product question: whether compatibility means proxying only familiar chat-style APIs, or also exposing specialized surfaces such as decisions, agents, search and tool execution.
Why it matters for gateways
Most AI gateway architecture began with a relatively simple assumption: normalize provider differences so developers can switch models behind one API. The Decisions API pushes against that assumption. It is not just a different model behind the same chat request. It is a specialized endpoint with a different promise: low-latency typed output.
That affects routing, observability and billing. A classification call that returns a small typed object should not necessarily be analyzed beside a long-form generation request, even if both use a GPT-6-family model. Latency targets, success criteria and cost attribution can all be different. A fraud-screening decision, an image moderation label and a sales-lead routing outcome may have more in common with each other than with a conversational assistant turn.
For an AI API control plane for developer teams, that means the catalog needs a clearer distinction between model identity and endpoint capability. A row that says “GPT-6 Luna” is not enough if one version is exposed through a general Responses flow and another through a typed Decisions endpoint. Access controls may also need to be endpoint-aware, because a team may want to permit low-risk classification workloads while restricting broader generation use.
Model Gate and similar OpenAI-compatible gateways have a direct product implication here: compatibility now has to cover workflow semantics, not just request authentication and model names. If customers expect one API key, one usage ledger and one set of team policies, decision calls need to be visible as their own workload class rather than hidden inside generic GPT usage.
Billing and analytics get more granular
The launch also adds pressure to billing systems that already struggle to make different AI workloads comparable. Token usage remains important, but a decision endpoint is often bought for throughput, latency and reliability. If the endpoint is marketed as much faster than the Responses API for the same class of task, customers will want analytics that show whether they are actually getting that operational benefit.
That creates a need for separate reporting dimensions: endpoint, decision schema, model, latency, error rate and cost. In a platform built around unified AI API billing, those fields help explain why two calls to the same underlying model family may appear under different product categories or price bands.
Reseller and partner platforms face a related issue. If they expose the Decisions API to their own customers, they may need separate documentation, margin rules and usage exports for decision workloads. Treating every Luna call as ordinary text generation could blur margins and make customer invoices harder to defend.
What remains uncertain
The beta label leaves several open questions. OpenAI has not yet turned the Decisions API into a broad multi-model surface, and GPT-6 Luna is the only model listed for now. It is also unclear how quickly SDKs, OpenAI-compatible gateways and observability tools will converge on common naming and tracing conventions for decision calls.
There is also a product-design question for developers. Some applications will benefit from a dedicated typed endpoint immediately, especially where prompt parsing has been brittle or slow. Others may prefer to keep classification inside a broader Responses workflow if they depend on multi-step reasoning, tool calls or shared context.
The safest reading is that OpenAI is carving out another specialized API category. The broader gateway market has been moving in the same direction: search, agents, hosted tools, audio and now typed decisions are becoming separately metered surfaces. For developers, that can mean better fit and lower latency. For platform teams, it means the gateway layer has more to understand than a base URL and a model string.