Models & Providers

Your OpenClaw agent is model-agnostic. The same agent configuration, skills, and tools work across any supported model — swap the brain, keep everything else. This guide helps you decide which brain to pick.

How OpenClaw Handles Models

OpenClaw talks to LLM providers through three protocol adapters:

• openai-completions — covers roughly 80% of providers (OpenAI, DeepSeek, Grok, GLM, Qwen, Step, and any OpenAI-compatible API)
• anthropic-messages — for Claude models
• google-generative-ai — for Gemini models

Because most providers follow the OpenAI-compatible format, adding a new one is a configuration change — a base URL and an API key — not a code change. This is why OpenClaw can support 20+ providers without a bloated codebase.

Available Models

The table below lists the models currently available on Claw Buff. Quality index is based on public benchmarks (higher is better, scale 0–4). Speed is measured in tokens per second. Blended cost is the weighted average of input and output token pricing.

Which Model Should You Pick?

There is no single best model. The right choice depends on what you need your agent to do.

Best for general agent work

Claude Sonnet 4.6 — It has the strongest tool-calling accuracy and the best resistance to prompt injection among current models. If your agent calls tools, browses the web, or handles untrusted input, Sonnet is the safest default.

Highest raw intelligence

GPT-5.4 or Gemini 3.1 Pro — Both score idx 57 on quality benchmarks. GPT-5.4 tends to be more creative in open-ended tasks. Gemini 3.1 Pro is significantly faster (125 vs 77 t/s) and cheaper ($4.50 vs $5.63/M). Gemini 3.1 Pro is still labeled "preview" by Google, so expect occasional rough edges.

Best value

GLM-5 — Quality tier 4 at just $1.55/M tokens. It delivers benchmark scores close to Sonnet and GPT-5.4 at roughly a quarter of the price. A strong choice when you need good quality without the premium cost.

Best speed

Grok 4.20 Beta (227 t/s) or Gemini 3 Flash (166 t/s) — For interactive agents where response latency matters more than peak quality. Grok is still in beta and less battle-tested for complex agent workflows. Gemini 3 Flash trades quality for speed — it's tier 3, not tier 4.

Budget and high-volume

DeepSeek V3.2 ($0.29/M) or Step 3.5 Flash ($0.15/M) — When you're processing large volumes or experimenting and want to keep costs minimal. Step 3.5 Flash is the cheapest option available but is tier 2 quality — fine for simple Q&A, less reliable for multi-step tool-calling chains.

Premium and complex reasoning

Claude Opus 4.6 — The most expensive model on the list ($10.00/M) but excels at deep multi-step reasoning, nuanced analysis, and tasks that require careful deliberation. Use it when quality justifies the cost — legal analysis, code architecture decisions, or research synthesis.

Known Limitations by Provider

Every provider has trade-offs. Here's what to watch for:

• DeepSeek — The slowest model at 25 t/s. Quality can degrade on multi-step reasoning chains. Frequently hits rate limits during peak hours.
• OpenAI — Tool calling is less consistent than Claude's. Context window is 128K tokens vs Anthropic's 200K, which matters for agents with large memory or many tools.
• Google Gemini — Gemini 3.1 Pro is still in "preview" status. Gemini 3 Flash sacrifices quality for speed.
• xAI Grok — Still labeled "Beta." Less community testing for complex agent workflows compared to Claude or GPT.
• Z AI GLM-5 Turbo — Despite the name, throughput is only 3 t/s. The non-Turbo GLM-5 is much faster.

Cost on Claw Buff

Claw Buff uses a credit system. You purchase credits and spend them as your agent uses tokens. The exchange rate is straightforward: 1 cent = 10 credits. Your credit balance is visible in the Console, and you can track usage per model.

If you prefer to pay providers directly, the Hobby plan supports BYOK (Bring Your Own Key). You supply your own API key for any supported provider, and Claw Buff only charges the subscription fee — no markup on token usage. API keys are encrypted with AES-256-GCM at rest.

Model Fallback

OpenClaw supports automatic model fallback. You define a primary model and a list of fallbacks. If the primary fails — due to authentication errors, rate limits (HTTP 429), or timeouts — the agent automatically retries with the next model in the chain.

On Claw Buff, fallback chains are configured automatically based on your plan. On self-hosted OpenClaw, you configure them in openclaw.json:

{
  llm: {
    provider: "anthropic",
    model: "claude-sonnet-4.6",
    fallbacks: [
      { provider: "openai", model: "gpt-5.4" },
      { provider: "google", model: "gemini-3.1-pro" }
    ]
  }
}

Local Models with Ollama

If you run OpenClaw self-hosted, you can use local models through Ollama. OpenClaw connects to Ollama via the openai-completions adapter (Ollama exposes an OpenAI-compatible API at localhost:11434).

Popular local models include Llama 3, Mistral, Qwen 2.5, Phi-4, DeepSeek, and Gemma.

Pros:

• Complete data privacy — nothing leaves your machine
• No API costs after hardware investment
• No rate limits

Cons:

• Requires 16–32 GB of RAM depending on model size
• Slower than cloud APIs on consumer hardware
• Tool calling is weaker on most local models, which limits agent capabilities

Switching Models

On Claw Buff, open the Console, go to Settings > Model, and select a new model. The change takes effect on the next message — no restart needed.

On self-hosted OpenClaw, edit the llm section of your openclaw.json and restart the Gateway. Your conversation history, skills, and tools are preserved across model changes.