Model Configuration & Switching: Let OpenClaw Auto-Select the Best AI Model

One Model to Rule Them All? Not Quite

You might be using GPT-4o (or Gemini Flash) for everything. Most of the time, that’s fine.

But consider these scenarios:

Scenario	Using GPT-4o	Better Choice
Simple classification, format conversion	Overkill, wasting Tokens	GPT-4o mini (20x cheaper)
Complex mathematical reasoning	Sometimes gets it wrong	o1 (specialized for reasoning)
Very long documents (100K+ words)	Hits context limits	Gemini 1.5 Pro (1M Token window)
API goes down	Entire system stops	Auto-switch to Claude (Fallback)

Model switching isn’t an advanced feature — it’s the basics of saving money and staying reliable.

---

Provider Setup

What Is a Provider?

Provider = the “vendor” of AI models.

OpenAI    → GPT-4o, GPT-4o mini, o1
Google    → Gemini 2.0 Flash, Gemini 1.5 Pro
Anthropic → Claude 3.5 Sonnet, Claude 3 Haiku

Set Up API Keys

Configure each Provider’s Key in OpenClaw’s .env file:

# Set at least one
OPENAI_API_KEY=sk-xxxxxxxxxxxxxxxx
GOOGLE_API_KEY=AIzaSyxxxxxxxxxxxxxxxxxx
ANTHROPIC_API_KEY=sk-ant-xxxxxxxxxxxxxxxx

Don’t have an API Key yet? Check the AI Model API Key Guide first — Google AI Studio has a free tier, perfect for beginners.

Note: The key you enter here is the AI model Key (for calling LLMs), not the Google Drive/Gmail OAuth Key — those are different.

Test Connections

# Test all configured Providers
openclaw provider test

# Example output:
# ✅ OpenAI: Connected (GPT-4o, GPT-4o-mini)
# ✅ Google: Connected (Gemini-2.0-Flash, Gemini-1.5-Pro)
# ❌ Anthropic: No API key configured

---

Model Routing: Auto-Select the Best Model

Core Concept

Model routing = an AI version of “smart call forwarding.” Based on the task characteristics, it automatically assigns the most suitable model.

Simple Q&A        → GPT-4o mini (cheap & fast)
Writing/Analysis  → Claude Sonnet (best quality)
Long documents    → Gemini Pro (ultra-long context)
Math reasoning    → o1 (specialized reasoning)

How to Configure

In config.yaml:

models:
  # Default model (used when nothing else is specified)
  default: gpt-4o-mini

  # Model routing rules
  routing:
    # High-quality tasks
    - pattern: "write|create|compose|analyze|report"
      model: gpt-4o
      reason: "Requires higher generation quality"

    # Reasoning tasks
    - pattern: "calculate|prove|math|reasoning"
      model: o1
      reason: "Complex reasoning needs a specialized model"

    # Long document tasks
    - pattern: "summarize_long|full_text|long_form"
      model: gemini-1.5-pro
      reason: "Requires ultra-long context window"

  # Everything else
  fallback: gpt-4o-mini

Specify a Model in a Skill

You can also specify a model for each step within a Skill:

name: weekly-report
steps:
  # Step 1: Gather data (use cheap model)
  - action: llm_generate
    config:
      model: gpt-4o-mini
    input:
      prompt: "Extract key facts from the following news..."

  # Step 2: Write analysis (use premium model)
  - action: llm_generate
    config:
      model: gpt-4o
    input:
      prompt: "Based on the following facts, write an in-depth analysis report..."

  # Step 3: Translate to English (cheap model is enough)
  - action: llm_generate
    config:
      model: gpt-4o-mini
    input:
      prompt: "Translate the following report into English..."

This way, for one report: Steps 1 and 3 use the cheap model, and only Step 2 uses the premium one. Over 60% cheaper than using GPT-4o for the entire process.

---

Fallback Mechanism: Don’t Fear Downtime

What Is Fallback?

API services occasionally go down (rate limits, server maintenance…). The Fallback mechanism lets OpenClaw automatically switch to a backup model so your work isn’t interrupted.

Configure Fallback

models:
  default: gpt-4o

  fallback_chain:
    - gpt-4o           # First choice
    - claude-3.5-sonnet # If OpenAI is down, use Claude
    - gemini-2.0-flash  # If Claude is also down, use Gemini
    - gpt-4o-mini       # Last resort

  # When to trigger Fallback
  fallback_triggers:
    - error: rate_limit     # API rate limiting
    - error: server_error   # Server error
    - error: timeout        # Timeout (>30 seconds)
    - error: context_length # Context limit exceeded

How It Works in Practice

Trying GPT-4o... → 429 Rate Limit
Auto-switching: Trying Claude 3.5 Sonnet... → ✅ Success
Work continues, user notices nothing

Fallback is the biggest benefit of having multiple Providers. With only one Provider, you don’t have this safety net.

---

Model Comparison: Which Model for Which Scenario?

Quick Selection Guide

💰 Most affordable:    GPT-4o mini / Gemini Flash
📝 Best writing:       Claude 3.5 Sonnet
🧮 Best reasoning:     o1
📚 Longest documents:  Gemini 1.5 Pro (1M tokens)
⚡ Fastest response:   Gemini 2.0 Flash
🎨 Best multimodal:    GPT-4o / Gemini Pro

Detailed Comparison

Feature	GPT-4o	GPT-4o mini	Claude Sonnet	Gemini Flash	o1
Speed	Medium	Fast	Medium	Fastest	Slow
Quality	High	Medium	Highest (writing)	Medium	Highest (reasoning)
Price	Medium	Lowest	Medium-high	Free/Low	High
Context Window	128K	128K	200K	1M	200K
Tool Use	Best	Good	Good	Good	Limited

My Recommended Combos

Beginner (zero cost):

default: gemini-2.0-flash

Daily use (low cost):

default: gpt-4o-mini
fallback: gemini-2.0-flash

Professional work (quality first):

default: gpt-4o
routing:
  - pattern: "write|analyze"
    model: claude-3.5-sonnet
  - pattern: "math|reasoning"
    model: o1
fallback: gpt-4o-mini

Not sure how to estimate costs? See Token Economics for detailed cost calculations.

---

Advanced: Cost Budget Controls

Set Daily/Monthly Budget Limits

budget:
  daily_limit: 1.00    # Max $1 USD per day
  monthly_limit: 20.00 # Max $20 USD per month

  # What happens when over budget
  over_budget_action: downgrade  # Auto-downgrade to cheapest model
  # Other options: block (stop entirely), warn (just warn)

Track Usage

# View today's usage
openclaw usage today

# Output:
# 📊 Today's Usage Stats
# ├─ GPT-4o:     2,340 input / 1,560 output = $0.021
# ├─ GPT-4o-mini: 15,200 input / 8,900 output = $0.008
# └─ Total: $0.029 / $1.00 daily limit (2.9%)

# View monthly report
openclaw usage monthly

---

FAQ

I set up multiple Providers but only one gets used?

Check the routing rules in config.yaml. Without routing configured, only the default model is used.

Model response quality is inconsistent?

Fallback might be triggering and switching to a lower-quality model. Check the logs:

openclaw logs --filter model
# Shows which model was actually used for each request

Can anyone see my API Key?

No. API Keys are stored in the .env file and only used on the local server side. The frontend (browser) never sees the Key.

If you’re using cloud deployment, Keys are stored in environment variables and are equally secure.

---

Next Steps

Now that your models are configured:

• 🧩 Use different models for different steps in your Skills
• 🤖 Let Agents auto-select models to complete tasks
• 💬 Master Prompts to get great results from cheap models
• 💰 Calculate your Token costs precisely
• 🔗 Connect more external tools with MCP