Choose a Model for OpenClaw

The right model is not simply the model with the best chat score. An OpenClaw model must follow instructions reliably, produce valid tool calls, recover from tool errors, and fit your latency and cost limits.

Provider configuration

The onboarding flow configures authentication and a primary model for common providers:

openclaw onboard

Then inspect the effective state:

openclaw models status
openclaw models list

Exact model references

OpenClaw model references use provider/model. Set a primary model with:

openclaw models set <provider/model>

Avoid copying a display name from a provider dashboard. Use the exact reference printed by openclaw models list --provider <provider>.

Choose by workload

Older or very small models can sound fluent while producing malformed parameters or looping after a failed tool. Test the actions you actually need, not only a short chat prompt.

Verify the selected model

Run a harmless tool task and inspect status again. If you configure fallbacks, remember that an explicit session model selection is intentionally strict and may fail visibly instead of moving through the configured fallback chain.

Expected result

openclaw models status shows usable authentication and the intended primary reference. A test turn completes without a provider auth error or a model-policy rejection.

Continue by setup

• OpenAI: subscription OAuth or direct API-key billing.
• Anthropic: API key or Claude CLI on the Gateway host.
• Local models: hardware fit, privacy boundary, and Ollama.
• Tool-calling test: a repeatable promotion checklist.
• Coding models: repository edits, commands, and verification.