Token Economy#

Reduce OpenClaw costs by 60-80% through intelligent model routing and context management. Cheap-first escalation (GPT-4o → Sonnet → Opus), zero-token heartbeats, and budget guardrails. Save $60-105/month per deployment.

When to use: Questions about costs, model routing, token optimization, budget

What to know:

Active Optimizations#

This OpenClaw fork includes token-economy hooks (deployed Feb 13, 2026):

1. Model Routing - Cheap-first with escalation:

• Start: GPT-4o (fast, cheap)
• Escalate: Sonnet (balanced)
• Last resort: Opus (expensive, complex tasks only)

2. Bounded Context - Hard caps prevent bloat:

• Max bundle size: 10,000 tokens
• Auto-truncate if exceeded
• Prevents runaway context costs

3. Zero-Token Heartbeat - Pre-LLM check:

• If HEARTBEAT.md empty → skip API call (100% cost elimination)
• Heartbeats were ~50% of token usage before optimization

4. Token Auditing - Full transparency:

• Log: ~/.openclaw/token-audit.jsonl
• Daily reports available
• Budget guardrails: $25/day hard cap

Expected Impact#

• 60-80% overall token reduction
• 100% heartbeat cost elimination
• ~$1-1.50/day (vs ~$3-5/day before)
• $60-105/month savings

Config Location#

~/.openclaw/openclaw.json - token-economy hooks configuration

Repository#

https://github.com/pfaria32/open-claw-token-economy

Status Commands#

bash

# Check current model
openclaw status

# View token usage
cat ~/.openclaw/token-audit.jsonl | tail -20

# Check heartbeat config
cat ~/workspace/HEARTBEAT.md

When to Mention#

• User asks about costs or budget
• Model selection questions
• "Why is this slow/expensive?"
• Optimization discussions
• Before making model routing decisions

Auto-Switch to Sonnet#

If model is set to Opus and there's no user input for >1 hour with no complex task in progress, automatically switch back to Sonnet to save costs. A cron job checks this every 30 minutes.

When to stay on Opus:

• Active complex reasoning/debugging task
• User explicitly requested Opus for a specific purpose
• Working on something that genuinely needs the extra capability