Documentation
Model selection and automatic routing
Understand explicit model selection, experimental routing and which features are available today.
Use an explicit model today
The public API requires an enabled, concrete model ID from live models. Both Chat Completions and Messages use the model you specify. They do not silently choose a different model.
Check the model's capability category and pricing before sending a request. Image and speech models belong to their respective endpoints, not either chat format.
Automatic routing is experimental
The published routing library is experimental. New local auto, eco and premium planning work has been implemented but is not part of the published 0.2.0 CLI/SDK release. Do not send these aliases to the public API or assume a current host adapter supports them. Updated installation instructions will accompany a separately verified release.
The new planner filters models by capabilities, context/output limits, the owner's allowed models, health and budget, then compares bounded request-specific quotes. It selects one concrete model before the ordinary x402 payment flow. It does not widen wallet authority, automatically retry an uncertain payment, or guarantee a quality or cost advantage for every task.
Provider expansion
The September 9 release includes 93 Venice text models and seven Gemini text models on Vertex, alongside the existing Gemini image and ElevenLabs speech models. Venice uses explicit venice/ model IDs on both text endpoints. Google models remain on Vertex; Venice encrypted-only models are not supported.
Ten randomly selected models passed live provider inference checks. All 93 Venice models passed unsigned API challenge checks in both request formats. This is sampled compatibility evidence, not individual live inference certification or a paid customer test for every model. Only use IDs and capabilities returned by the live catalog; see text generation for limits.
Caching is a different feature
Local response caching reuses an eligible identical answer inside the same authorized local session. A cache hit performs no new paid request. Provider prompt-prefix caching, when supported and accounted for, instead reduces provider work on repeated input; it does not reuse the completed answer or imply a lower already-signed exact payment.
Current OpenClaw and Hermes turns retain explicit recovery keys and bypass response caching.