OpenRouter Auto Router vs sume/auto: who picks the model?

What does an auto router do on each platform?

Both pick a model for you, but they work on different media and disclose different things. OpenRouter's openrouter/auto selects among text models for a messages request. Sume's sume/auto on POST /v1/videos picks a video family for a generation request, and the response only ever says sume/auto.

How does the OpenRouter Auto Router decide?

From its documentation: model ids openrouter/auto and openrouter/auto-beta. The router classifies the task into roughly 30 fine-grained types, then ranks candidates by real-world spend share from trailing 7-day OpenRouter community data. It applies cost-tier filters (low through max, which filter candidates rather than set ceilings), honors account restrictions, and has fallback models. provider.max_price still applies, patterns such as anthropic/* can allow or exclude models, and session stickiness keeps a model across turns.

You pay the standard rate of whichever model is selected, with no additional fee for the router. It requires the messages format rather than prompt.

How does sume/auto decide, and what do you see?

Per Video generation, sume/auto is a Sume-only addition for when you do not want to pin a family. Resolution is a pure function of the normalized request and the catalog version, so an idempotent replay prices and routes identically. Sume does not disclose which family served the request and says not to infer it from the output.

Pricing is the other difference. Sume reserves workspace USD balance on submit at provider list x 1.25 for every model, and usage.cost is the Sume billable amount. OpenRouter states no router fee; Sume has no router fee line, but the multiplier applies to all models, pinned or auto. To pin a family instead, send a catalog id such as seedance-2.5 from GET /v1/videos/models.

When should you pin a model instead?

Pin when the output must be reproducible by family, when you need a specific limit (for example seedance-2.5 accepts 4 to 30 seconds while most catalog models top out at 15), or when your reviewers compare models. Use auto when the brief is simple and you care about getting a clip, not about which engine made it.

• Read supported_durations and supported_resolutions from GET /v1/videos/models before pinning.
• Send an Idempotency-Key on submit; a replay returns the original job.
• Log the job id, not an inferred model, because Sume will not tell you the family under auto.