LLM Provider Comparison

OpenClaw supports 12+ LLM providers out of the box. This reference compares them all so you can choose the right one for your use case and budget.

Provider Overview#

| Provider | Type | Models | Tool Calling | Pricing | Best For | |----------|------|--------|-------------|---------|----------| | Anthropic | Cloud | Claude Opus, Sonnet, Haiku | Excellent | $0.25-75/MTok | Complex reasoning, tool use | | OpenAI | Cloud | GPT-5, GPT-5.1-Codex, o3 | Excellent | $2-60/MTok | Speed, coding, wide compatibility | | Google | Cloud | Gemini 2 Pro, Flash | Good | $0.075-20/MTok | Large context, multimodal | | OpenRouter | Cloud | 200+ models | Varies | Varies | Model variety, experimentation | | Ollama | Local | Llama 3, Mixtral, Qwen, etc. | Good | Free | Privacy, no API costs | | vLLM | Local | Any HuggingFace model | Good | Free | High-throughput inference | | Amazon Bedrock | Cloud | Claude, Llama, etc. | Good | Varies | AWS integration | | Moonshot AI | Cloud | Kimi | Basic | Low | Chinese language tasks | | MiniMax | Cloud | Various | Basic | Low | Specialized tasks | | LM Studio | Local | Any GGUF model | Varies | Free | Desktop users, easy UI |

Detailed Provider Analysis#

Anthropic Claude (Recommended)#

Claude is the recommended provider for OpenClaw because it has the strongest reasoning and tool-use capabilities in 2026.

Model tiers:

• Claude Opus -- Best for complex, multi-step reasoning tasks. Use for research, analysis, and architecture decisions. Most expensive.
• Claude Sonnet -- Best balance of quality and cost. Recommended as default for most daily tasks.
• Claude Haiku -- Fastest and cheapest. Good for simple queries, quick lookups, and high-volume tasks.

Why Claude for OpenClaw:

• Superior tool calling reliability -- fewer hallucinated tool invocations
• Best understanding of complex multi-step instructions
• Strong safety guardrails against prompt injection
• Excellent at following agent prompt guidelines (AGENT.md files)

OpenAI GPT#

Model tiers:

• GPT-5 -- Flagship model. Strong general-purpose performance.
• GPT-5.1-Codex -- Optimized for code generation and tool use.
• o3 -- Reasoning model for complex problem-solving.

When to choose OpenAI:

• You need fast response times (GPT models tend to be faster)
• Your workflows are code-heavy (Codex models excel here)
• You are already in the OpenAI ecosystem

Ollama (Local Models)#

Ollama is the go-to choice for running models locally. Zero API costs, full privacy.

Best local models for OpenClaw (March 2026):

| Model | VRAM Required | Tool Calling | Quality | |-------|--------------|-------------|---------| | Llama 3 70B | 40 GB | Good | High | | Llama 3 8B | 8 GB | Moderate | Medium | | Mixtral 8x7B | 32 GB | Good | High | | Qwen 2.5 72B | 40 GB | Good | High | | Phi-3 Medium | 8 GB | Basic | Medium |

**Hardware requirements:** - Minimum: 8 GB VRAM for small models (8B parameters) - Recommended: 24+ GB VRAM for quality models (70B parameters) - Apple Silicon Macs work well with Ollama using unified memory

OpenRouter#

OpenRouter acts as a gateway to 200+ models from various providers through a single API key. Useful for experimentation and fallback routing.

Why use OpenRouter:

• Try different models without managing multiple API keys
• Automatic routing to the cheapest available model
• Good for fallback chains

Google Gemini#

Key advantage: Massive context windows (up to 2M tokens with Gemini 2 Pro). If your workflows involve processing large documents, codebases, or long conversation histories, Gemini's context window is unmatched.

Multi-Model Strategies#

Fallback Routing#

{
  "models": {
    "defaults": {
      "provider": "anthropic",
      "model": "claude-sonnet-4-20250514"
    },
    "fallbacks": [
      { "provider": "openai", "model": "gpt-5" },
      { "provider": "ollama", "model": "llama3:70b" }
    ]
  }
}

If Claude hits a rate limit, OpenClaw automatically falls back to GPT-5, then to Ollama.

Per-Channel Routing#

Route different channels to different models based on importance and cost:

• WhatsApp (personal): Claude Sonnet (best quality for important conversations)
• Discord (casual): Ollama Llama 3 (free, good enough for casual chat)
• Slack (work): GPT-5 (fast responses for team interactions)

Task-Based Routing#

Route by task complexity:

• Simple queries: Claude Haiku or Llama 3 8B (fast, cheap)
• Email triage: Claude Sonnet (good reasoning at moderate cost)
• Research and analysis: Claude Opus (best reasoning, worth the cost)
• Code generation: GPT-5.1-Codex (optimized for code)

Cost Optimization#

| Strategy | Savings | Trade-off | |----------|---------|-----------| | Use Haiku/small models for simple tasks | 80-95% | Lower quality on complex tasks | | Set daily spend limits | Predictable | May hit limits on busy days | | Use Ollama for non-critical channels | 100% | Requires GPU hardware | | Cache common responses | 30-50% | Stale data risk | | Use OpenRouter's cheapest routing | 20-40% | Variable quality |

# Set a daily spending limit
openclaw config set llm.limits.dailySpend 10

Security Considerations#

• Model choice affects security: Older and smaller models are significantly less robust against prompt injection
• Always use the latest generation of instruction-hardened models for tool-enabled agents
• Local models keep your data completely private but may have weaker safety guardrails
• API keys should be stored in environment variables, never in config files directly

Next Steps#

• LLM Quickstart -- Configure your first provider
• MCP Skills -- Give your agent more abilities
• Cost management tips -- Optimize your spending