OpenAI Decisions API vs Sume auto routing: different jobs

The OpenAI Decisions API, a limited preview announced at DevDay 2026, lets you define questions with predetermined answers so a model can classify content or route a request. Sume's auto routing is a different layer: when you omit the model on a hosted MCP generate_image or generate_video call, Sume routes to sume/auto unless the user named a model family. Decisions picks your branch; Sume picks the media model. They compose without overlapping.

What OpenAI describes

According to the DevDay announcements thread, the Decisions API is in limited preview, powered by Luna, and covers content classification and request routing with answer sets you predefine. I found no further detail in the sources I could read, so I do not describe its pricing, limits or response shape. The OpenAI changelog is the place to check when it graduates.

What Sume auto routing is

The hosted MCP docs say paid generation tools such as generate_image and generate_video route to sume/auto when payload.model is omitted, unless the user named a family. That choice is about which media model renders the job. It does not decide whether a request is a refund question, a product shot or a policy violation. For the list of models a router exposes, see the video router page and the tool list in MCP tools and gates.

A combined flow

Use the classifier first, and keep the paid step explicit:

• Classify the incoming request into one of a few answers: new clip, edit existing clip, question, out of scope.
• Only the first two branches reach a Sume tool. Everything else answers in text and spends nothing.
• On the media branch, call generation_admission_preview or dry_run=true, then submit with an idempotency_key.
• Keep the model unset unless a person asked for a family, so sume/auto can route.

Why not let one model do both

A single agent can classify and render, and Sume's own tools work without a classifier. A separate decision step helps when you want a cheap, auditable gate before any paid call, or when different answers go to different systems entirely. It adds a dependency on a preview feature, so keep a plain fallback, such as a keyword rule, that routes to read-only behavior if the preview is unavailable to your account.

Whichever you pick, store the answer next to the Sume job id so you can explain later why a render started.