Models & Routing
Model catalog, dynamic routing, state affinity and caching behavior.
Model catalog
GET /v1/models returns ID, display name, context length and capabilities such as streaming, vision and tools where supported.
Model onboarding
For new OpenAI/Anthropic models the operational connection target is up to 24 hours after production API availability where technically feasible. For an off-catalog model: feasibility review up to 48 hours; if feasible, target connection within 7 calendar days. These are operational targets, not financial SLA.
Dynamic routing
For independent requests the upstream is selected dynamically.
State affinity
Related requests preserve upstream affinity. If continuation requires the original upstream but affinity is unknown, HTTP 409 state_affinity_unknown is returned.
Caching
Caching is used where supported by the model/upstream. No cross-client prompt-content cache is used; sticky affinity preserves cache locality.