Reduce API Costs — Save Money on AI Usage

## Quick Wins (Immediate Savings)

### 1. Switch to Sonnet for Most Tasks

Claude 3.5 Sonnet is 5x cheaper than Opus and handles 90% of tasks equally well. Set it as your default:

# ~/.openclaw/config.yamlmodels:  primary: anthropic/claude-sonnet-4-20250514  fallbacks:    - anthropic/claude-3-5-haiku-20241022    - openai/gpt-4o-mini

Cost comparison per 1M tokens: - Opus: $15 output / $75 input - Sonnet: $3 output / $15 input - Haiku: $0.25 output / $1.25 input

Use Opus only when you explicitly need it (complex reasoning, long documents).

### 2. Enable Prompt Caching

Anthropic's prompt caching stores your system prompt and conversation context, charging only 10% for cached tokens:

models:  anthropic:    cacheControlTtl: 300  # Cache for 5 minutes

This alone can reduce costs by 50-80% for conversational use.

### 3. Configure Aggressive Context Pruning

Don't pay for old messages you don't need:

contextPruning:  mode: sliding  maxMessages: 20  # Keep only last 20 messages  maxTokens: 50000  # Cap at 50k tokens

For most conversations, 20 messages of context is plenty.

### 4. Use Haiku for Heartbeats & Cron

Automated tasks don't need the smartest model:

agents:  main:    heartbeat:      model: anthropic/claude-3-5-haiku-20241022

Haiku is 60x cheaper than Opus — perfect for scheduled checks.

## Advanced Strategies

### 5. Set Up Model Routing

Route different tasks to appropriate models:

# Use Haiku for simple queries, Sonnet for code, Opus for analysismodels:  primary: anthropic/claude-sonnet-4-20250514  routing:    simple: anthropic/claude-3-5-haiku-20241022    code: anthropic/claude-sonnet-4-20250514    analysis: anthropic/claude-opus-4-20250514

### 6. Use Local Models for Drafts

Run Ollama locally for first drafts, only use paid APIs for final output:

models:  local: ollama/qwen2.5:7b  primary: anthropic/claude-sonnet-4-20250514

Local models cost $0. Use them for brainstorming, then polish with Claude.

### 7. Truncate Tool Outputs

Large file reads and web scrapes bloat context. Limit tool output size:

tools:  read:    maxChars: 10000  # Cap file reads at 10k chars  web:    maxChars: 5000   # Cap web fetches

### 8. Monitor with Session Status

Check your token usage regularly:

# In chat/status# CLIopenclaw status --usage

This shows tokens used and estimated cost per session.

### 9. Set Budget Alerts

Configure alerts in your provider dashboard: - Anthropic Console: console.anthropic.com → Usage → Set alerts - OpenAI: platform.openai.com → Settings → Limits

Set alerts at 50% and 80% of your monthly budget.

### 10. Use Batch Processing for Bulk Work

If you're processing many items (emails, documents), batch them instead of one-by-one. Anthropic offers batch API with 50% discount.

## Cost Estimation Cheat Sheet

Typical monthly costs by usage pattern:

| Usage | Model | Est. Monthly Cost | |-------|-------|-------------------| | Light (10 msgs/day) | Haiku | $1-3 | | Medium (50 msgs/day) | Sonnet | $5-15 | | Heavy (200+ msgs/day) | Sonnet + caching | $15-40 | | Power user | Opus + Sonnet mix | $30-80 |

With caching enabled, expect 50-80% reduction from these estimates.

## Example: Optimized Config

# Optimized for costmodels:  primary: anthropic/claude-sonnet-4-20250514  fallbacks:    - anthropic/claude-3-5-haiku-20241022    - minimax/MiniMax-M2.1  anthropic:    cacheControlTtl: 300contextPruning:  mode: sliding  maxMessages: 25  maxTokens: 60000agents:  main:    heartbeat:      model: anthropic/claude-3-5-haiku-20241022tools:  read:    maxChars: 15000  web:    maxChars: 8000

This config uses Sonnet by default, Haiku for automated tasks, aggressive caching, and limits context size. Expected savings: 60-80% vs naive Opus-only config.