Validated Inference Capabilities

Every model on Inference Providers now advertises capabilities that have been probed rather than declared: tool calling, structured output, vision input, and streaming are each validated against the live provider deployment before they appear on the model card.

Validation reruns on a 72-hour cycle. A capability that stops responding correctly is withdrawn from the card instead of silently failing at request time, and routing stops sending traffic that depends on it to that deployment.

The practical effect is that a request built against a model card's declared capabilities will be accepted by whichever provider serves it. If you route by capability rather than by provider name, you no longer need per-provider special cases.