OpenClaw does not have a tiny fixed model list. That is the point. It can talk to first-party model providers, local runtimes, and any OpenAI-compatible endpoint you can put behind a URL.
That flexibility is useful, but it also makes the first setup confusing. Here is the practical list.
Best starting models
| Model | Provider | Why Start Here |
|---|---|---|
| MiniMax M2.5 | Haimaker | Cheap default for everyday agent work |
| Gemini 3 Flash | Google or Haimaker | Long context, strong enough for most coding |
| Grok 4.1 Fast | xAI or Haimaker | Huge context at a low price |
| DeepSeek V3.2 | DeepSeek or Haimaker | Cheap output tokens and decent coding |
| Claude Sonnet 4.6 | Anthropic | Reliable tool use and coding quality |
| Gemma 4 | Ollama | Local, private, no token bill |
If you are new, do not start by configuring ten models. Pick one cheap default and one stronger fallback. That is enough.
Provider support
Haimaker
Haimaker is the easiest first option if you want one API key across many models. It exposes an OpenAI-compatible API, so OpenClaw configuration is straightforward.
Use it for MiniMax, Qwen, Llama, DeepSeek, Grok, Gemini, and other models without juggling separate provider accounts.
Anthropic
Claude models are still the premium coding-agent choice. Use Claude Sonnet as a serious default if cost is not your main constraint. Use Opus for hard debugging, not every prompt.
OpenAI
OpenAI models are good general-purpose choices with reliable API behavior. GPT-5.4 Mini is a reasonable value pick. GPT-5.4 Pro is too expensive to leave as your default unless someone else is paying.
Google Gemini
Gemini is the long-context family. Use Flash for most work, Pro when you need better reasoning, and 2.0 Flash when the free tier matters.
xAI Grok
Grok 4.1 Fast is the value pick. Grok Code Fast is better for code-heavy sessions. Grok 4.20 is the reasoning fallback.
DeepSeek
DeepSeek V3.2 is a strong low-cost option, especially when output volume matters. Keep a fallback because DeepSeek availability can be uneven.
Ollama
Ollama gives OpenClaw local models. Good picks include Gemma 4, Qwen3.5, and Llama 3.3. The price is zero per token, but the cost moves to your hardware and patience.
OpenAI-compatible providers
This is where OpenClaw gets flexible. If a provider supports the OpenAI chat completions format, you can usually add it.
Common examples:
- Haimaker
- Ollama
- OpenRouter
- Together
- Fireworks
- LM Studio
- vLLM
- Internal company gateways
The setup pattern is always the same:
{
models: {
providers: {
haimaker: {
baseUrl: "https://api.haimaker.ai/v1",
apiKey: "your-api-key",
api: "openai-completions",
models: [
{ id: "minimax/minimax-m2.5", name: "MiniMax M2.5" }
]
}
}
}
}
Then allowlist the fully qualified model name:
{
agents: {
defaults: {
models: {
"haimaker/minimax/minimax-m2.5": {}
}
}
}
}
That second step is where most setup errors happen.
Model list by job
Cheapest daily work
Use MiniMax M2.5, DeepSeek V3.2, Gemini Flash, or Grok 4.1 Fast.
Best coding quality
Use Claude Sonnet, Claude Opus, Gemini 3.1 Pro, or GPT-5.4.
Long context
Use Gemini, Grok 4.1 Fast, Grok 4.20, or Qwen models through Haimaker.
Local and private
Use Gemma 4, Qwen3.5, or Llama through Ollama.
The setup I would use
For a balanced OpenClaw config:
- Haimaker as the main provider
- MiniMax M2.5 as the default
- Gemini 3 Flash for long-context work
- Claude Sonnet or Gemini 3.1 Pro as the hard-task fallback
- Ollama + Gemma 4 if local privacy matters
That gives you a cheap baseline, a long-context option, and a stronger model when the cheap one stalls.