Models & Providers

Read this when you want to change which AI model your agent uses, switch providers, or run a local model.

Current model support

OpenClaw supports 35+ providers. Here are the most commonly used, followed by the full list.

Popular providers

All official providers

Amazon Bedrock, Anthropic, Claude Max API Proxy, Cloudflare AI Gateway, Deepgram (transcription), DeepSeek, GitHub Copilot, GLM, Google (Gemini), Groq, Hugging Face, Kilocode, LiteLLM, MiniMax, Mistral, Model Studio, Moonshot AI, NVIDIA, Ollama, OpenAI, OpenCode, OpenCode Go, OpenRouter, Perplexity, Qianfan, Qwen, SGLang, Synthetic, Together AI, Venice, Vercel AI Gateway, vLLM, Volcengine, xAI, Xiaomi, Z.AI

Changing your default model

In ~/.openclaw/openclaw.json, model references use the "provider/model" format:

{
  "ai": {
    "model": "anthropic/claude-sonnet-4-6",
    "apiKey": "YOUR_API_KEY"
  }
}

To switch to OpenAI:

{
  "ai": {
    "model": "openai/gpt-4o",
    "apiKey": "YOUR_OPENAI_API_KEY"
  }
}

Per-automation model overrides

You don't have to use the same model for everything. Set a cheaper model for routine tasks:

{
  "automations": [
    {
      "id": "morning-briefing",
      "model": "claude-haiku-4-5-20251001",
      "schedule": "0 7 * * *",
      "prompt": "..."
    }
  ]
}

Good defaults by task type:

Cost comparison

Approximate costs per 1,000 tokens (input/output combined, as of March 2026):

A typical conversation: 1,000-3,000 tokens. A morning briefing with skills: 2,000-5,000 tokens. A research task: 10,000-30,000 tokens.

At Haiku prices, you'd need to send 1,000 messages to spend $1. At Sonnet prices, about $0.01/message.

Multiple providers

You can configure multiple providers and reference them by name:

{
  "providers": {
    "fast": {
      "model": "anthropic/claude-haiku-4-5-20251001",
      "apiKey": "YOUR_KEY"
    },
    "smart": {
      "model": "anthropic/claude-sonnet-4-6",
      "apiKey": "YOUR_KEY"
    }
  },
  "ai": {
    "default": "fast"
  }
}

Then in automations or skill config, reference "provider": "smart" for tasks where you want the better model.

OpenAI-compatible API

The Gateway exposes OpenAI-compatible endpoints so external tools, IDE plugins, and RAG pipelines can use your OpenClaw agent as a drop-in OpenAI replacement:

Point any OpenAI-compatible client at http://localhost:18789 as the base URL. Explicit model overrides sent through /v1/chat/completions and /v1/responses are forwarded to the underlying provider, so clients that specify a model (e.g., a RAG pipeline requesting a specific embedding or completion model) work correctly.

Local models via Ollama

For offline use or privacy-sensitive tasks, run a local model with Ollama.

1. Install Ollama:

curl -fsSL https://ollama.ai/install.sh | sh

2. Pull a model:

ollama pull llama3.2
ollama pull mistral

3. Configure OpenClaw:

{
  "ai": {
    "model": "ollama/llama3.2",
    "baseUrl": "http://localhost:11434"
  }
}

Local models are slower and less capable for complex reasoning, but free and private. Good for simple automations and summaries.

Note: Local models don't support tool use as reliably as cloud models. Disable tools or use simple prompts if you hit issues.

Switching providers on the fly

You can ask your agent to use a different model mid-conversation:

"For this next task, use the smart model — I want a thorough analysis"

Add to SOUL.md:

When I say "use the fast model" switch to claude-haiku-4-5-20251001 for that response.
When I say "use the smart model" switch to claude-sonnet-4-6.

This requires model_override: true in your config:

{
  "ai": {
    "model_override": true
  }
}