Remote OpenClaw Blog
How to Use Free Models on OpenClaw via OpenRouter
7 min read ·
Remote OpenClaw Blog
7 min read ·
OpenRouter is a unified API gateway that provides access to 300+ language models from every major provider — OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek, and dozens of open-source model hosts — through a single API key. Instead of managing separate accounts with each provider, you use one OpenRouter key and switch between models by changing a single configuration field.
For OpenClaw operators, OpenRouter solves the multi-model problem. You can test GPT-5.4 in the morning, switch to Claude Opus at lunch, and run DeepSeek V3.2 overnight — all without changing your authentication, endpoint, or billing setup. It also provides automatic failover: if one provider has an outage, OpenRouter routes to an alternative backend.
The free tier is what makes OpenRouter particularly valuable for operators getting started. With 29 free models and 20 requests per minute, you can build, test, and run light production agents without spending a dollar on API costs.
OpenRouter's free tier works like this:
The 20 requests per minute limit translates to roughly 1,200 requests per hour or 28,800 per day. For a single-user agent or small team, this is a substantial amount of free compute.
Not all 29 free models are equally useful for agent workflows. Here are the best options ranked by agent performance:
| Model | OpenRouter ID | Best For | Context |
|---|---|---|---|
| GPT-OSS 20B | openai/gpt-oss-20b:free | Coding, general tasks | 128K |
| GPT-5.4-nano | openai/gpt-5.4-nano:free | Classification, routing | 128K |
| Llama 3.3 70B | meta-llama/llama-3.3-70b:free | General-purpose, writing | 128K |
| DeepSeek R1 Distill | deepseek/deepseek-r1-distill:free | Reasoning, math, analysis | 64K |
| Model | OpenRouter ID | Best For | Context |
|---|---|---|---|
| Qwen3 32B | qwen/qwen3-32b:free | Multilingual, reasoning | 32K |
| Gemma 3 27B | google/gemma-3-27b:free | Multimodal, general | 32K |
| Mistral Small | mistralai/mistral-small:free | Fast responses, European langs | 32K |
| Phi-4 14B | microsoft/phi-4:free | Reasoning, small footprint | 16K |
GPT-OSS 20B is the standout free model. Released by OpenAI under Apache 2.0, it matches o3-mini on coding benchmarks while being completely free on OpenRouter. For OpenClaw operators who want the strongest free coding agent, GPT-OSS 20B is the clear choice.
GPT-5.4-nano is free directly from OpenAI and also available free on OpenRouter. It is not powerful enough for complex agent tasks but excels as a router — analyzing incoming tasks and deciding which model should handle them.
Go to openrouter.ai and sign up with your email. No credit card required for free models.
Navigate to the API Keys section in your OpenRouter dashboard and create a new key. Copy it — you will need it for the OpenClaw configuration.
# In your OpenClaw config (e.g., ~/.openclaw/config.yaml)
llm:
provider: openrouter
model: openai/gpt-oss-20b:free # Best free coding model
api_key: your-openrouter-api-key
temperature: 0.7
max_tokens: 8192openclaw startThat is it. Your agent is now running on a free model with zero ongoing costs. The only limitation is the 20 requests per minute rate limit.
One of OpenRouter's strengths is instant model switching. To change models, update a single line in your config:
# Switch to Llama 3.3 70B (free)
model: meta-llama/llama-3.3-70b:free
# Switch to DeepSeek R1 Distill (free)
model: deepseek/deepseek-r1-distill:free
# Switch to GPT-5.4-nano (free)
model: openai/gpt-5.4-nano:free
# Switch to a paid model (requires credits)
model: anthropic/claude-opus-4-6-1mNo API key changes, no endpoint changes, no authentication updates. This makes it easy to test different models on the same workflow and find the best fit before committing to a paid model for production.
Marketplace
Free skills and AI personas for OpenClaw — browse the marketplace.
Browse the Marketplace →Here are proven strategies for running OpenClaw agents at zero cost:
# Primary: OpenRouter free model
llm:
provider: openrouter
model: openai/gpt-oss-20b:free
api_key: your-openrouter-api-key
# Fallback: Local Ollama (when rate-limited)
llm_fallback:
provider: ollama
model: qwen3:8b
base_url: http://localhost:11434Use OpenRouter for complex tasks (GPT-OSS 20B is stronger than local 8B models) and fall back to local Ollama when you hit the 20 req/min limit. This gives you unlimited capacity with the best model for each request.
# Router: Free nano classifies tasks
router:
model: openai/gpt-5.4-nano:free
# Complex tasks: Strong free model
complex:
model: openai/gpt-oss-20b:free
# Simple tasks: Fast local model
simple:
model: qwen3:8b # via OllamaUse the free GPT-5.4-nano to analyze each incoming task and route it to the appropriate model. Complex tasks go to GPT-OSS 20B; simple tasks go to your local Qwen3 8B. Total cost: zero.
Use free models for all development and testing. Once your agent workflow is validated, switch to a paid model (DeepSeek V3.2 at $0.028/$0.10) for production. This approach means you never pay for failed experiments or debugging sessions.
Consider adding credits to your OpenRouter account when:
The transition is seamless — just add credits and change the model ID. No other configuration changes needed.
Yes, genuinely free — no credit card required, no trial period, no hidden charges. OpenRouter offers 29 models at zero cost, funded by model providers who want distribution and usage data. The trade-off is rate limiting: free tier users get 20 requests per minute, which is enough for development, testing, and light production use but not for high-volume agent workflows.
GPT-OSS 20B (openai/gpt-oss-20b) is currently the strongest free model for coding tasks, matching o3-mini on code generation benchmarks. DeepSeek R1 Distill (deepseek/deepseek-r1-distill:free) is a close second with strong reasoning capabilities. For smaller, faster coding tasks, Qwen3 8B via Ollama locally may outperform both due to zero latency overhead.
For low-volume production (under 20 requests per minute), yes. The free tier rate limit of 20 req/min is sufficient for personal agents, small team deployments, and low-traffic applications. For anything above that volume, you will need to add credits to your OpenRouter account or use a direct API. There is no SLA on free models — availability and response times are best-effort.
Change the model field in your OpenClaw config. Free models on OpenRouter typically have ':free' appended to their model ID (e.g., 'meta-llama/llama-3.3-70b:free'). To switch to a paid version, remove the ':free' suffix. Your OpenRouter API key works for both free and paid models — the billing is per-model, so free models cost nothing even on a funded account.