I cut my OpenClaw API bill by 80% with one config change

Problem

• OpenClaw routes every request to your primary model by default. Trivial tasks (heartbeats, quick lookups, sub-agent work) use the same expensive model as complex reasoning, which wastes money.
• There is no built-in provider fallback by default, so provider rate limits can halt your agent.

Solution: model tiering + config change

Use different models for different task tiers and add cross-provider fallbacks to avoid outages.

Model tiers

• Frontier models (heavy reasoning)
  • Examples: Opus, GPT-5.2 — for architecture, large refactors, and other complex reasoning.
• Mid-tier models (daily work)
  • Examples: Sonnet, DeepSeek R1 — for code generation, research, content.
• Cheap / utility models (background, heartbeats)
  • Examples: Gemini Flash-Lite, DeepSeek V3.2, GLM 4.7 — for heartbeats, quick lookups, classification, background checks.

Fallbacks

• Add a fallback chain that spans providers (e.g., fall back to GPT-5.2 or another provider before falling to a same-provider mid-tier) to prevent total outages when a provider is rate-limited.

Model cost / performance examples

• Opus: ~ $30 per million tokens; ~50 tokens/sec.
• GPT-5.2: similar premium pricing.
• Sonnet / DeepSeek R1: mid-tier; cheaper with good reasoning (DeepSeek R1 cited at $2.74 per million tokens).
• Gemini 2.5 Flash-Lite: $0.50 per million tokens.
• DeepSeek V3.2: $0.53 per million tokens.
• Gemini 3 Flash: ~250 tokens/sec (faster than Opus).
• Cheap models can be ~60x cheaper than Opus for trivial requests.

Implementation options

• Manual configuration (recommended)
  • More control: explicitly map tasks to models and define fallbacks.
• OpenRouter auto-router
  • Automatic routing based on prompt complexity; less control but minimal setup.

Practical how-to (steps shown)

1. Edit the OpenClaw config file: ~/.openclaw/openclaw.json.
2. Define model sections for heartbeats, sub-agents, vision tasks, and fallback chains.
3. Create aliases for model shortcuts (e.g., opus, sonnet, flash, ds).
4. Restart the OpenClaw gateway.
5. Switch models on the fly with commands like /model sonnet, /model opus, or use /models to list providers.
6. Test to ensure heartbeats and background tasks now use cheaper models while primary agent tasks still use the frontier model.

Config highlights (example)

• Heartbeats → Gemini 2.5 Flash-Lite (~$0.50/m) instead of Opus.
• Sub-agents → DeepSeek R1 (~$2.74/m) instead of Opus.
• Primary tasks remain on Opus / GPT-5.2.
• First fallback uses a different provider (e.g., GPT-5.2) to avoid provider-wide rate limits.

Savings examples and tools

• Interactive calculator: https://calculator.vlvt.sh
• Example savings (monthly):
  • Light user: ~$200 → ~$70 (≈65% less).
  • Power user: ~$943 → ~$347 (≈62% less; ≈$600 saved/month).
  • Heavy user: ~$3,000 → ~$1,000 (≈62% less; ≈$1,700 saved/month).
• You can plug in different primary/heartbeat/sub-agent models to see custom savings and copy config snippets.

Why not rely on free models

• Free tiers commonly have aggressive rate limits, slow performance, and can disappear unexpectedly.
• The recommendation is to use near-free paid models (tens of cents per million tokens) for production reliability.

Links and resources mentioned

• Config examples and openclaw.json snippets: referenced in the video description.
• Savings calculator: https://calculator.vlvt.sh
• Model/provider lists and aliases: available in the video description.

Actionable checklist to apply now

• Add model-tiering to ~/.openclaw/openclaw.json:
  • Heartbeat → cheap model
  • Sub-agents → mid-tier
  • Primary → frontier
• Add a cross-provider fallback chain.
• Create model aliases for quick switching.
• Restart the gateway and test switching with /model commands.
• Use the calculator (calculator.vlvt.sh) to estimate savings for your usage.

Main speaker and sources

• Video presenter / author (demonstrates the config and calculator).
• Providers/models referenced: Anthropic (Opus, Sonnet), OpenAI (GPT-5.2), Google (Gemini Flash variants), OpenRouter, DeepSeek (R1, V3.2), GLM 4.7, Kimi K2.5.