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

📖 Description

If you’re running OpenClaw, there’s a good chance you’re burning money right now without realizing it.

In this video, VelvetShark shows you how to set up multi-model routing to cut your API costs by 50-80% - without losing quality on the tasks that actually matter.

Channel: VelvetShark URL: https://www.youtube.com/watch?v=fkT41ooKBuY

🎯 Learning Objectives

By the end of this video, you will understand:

Why OpenClaw’s default configuration costs more than it should
How to implement model tiering for different task types
The exact configuration changes to cut costs by 50-80%
How to use the /model command for dynamic cost control
Price differences between frontier, mid-tier, and cheap models
When to use free vs. paid tiers for production work

📋 Curriculum/Contents

0:00 - The problem: you’re burning money 0:39 - Why OpenClaw costs so much by default

Heartbeats use Opus ($30/M tokens)
Sub-agents all use primary model
Simple queries hit expensive models
No automatic fallback when rate limited

1:26 - The fix: model tiering

Complex reasoning → Frontier models (Opus, GPT-5.2)
Daily work → Mid-tier (Sonnet, DeepSeek R1)
Simple tasks → Cheapest models (Gemini Flash-Lite, DeepSeek V3.2)

2:22 - Model price comparison

Opus: $30/M tokens
GPT-5.2: $11.25/M tokens
DeepSeek R1: $2.74/M tokens
Gemini 2.5 Flash-Lite: $0.50/M tokens
60x price difference between cheapest and most expensive

3:03 - Manual config vs auto routing

Manual: More control, requires setup
OpenRouter auto-router: Less control, no configuration

3:34 - Copy-paste config walkthrough

Heartbeat configuration: Gemini 2.5 Flash-Lite
Sub-agents: DeepSeek R1
Fallback chain: GPT-5.2 first (different provider)
Model aliases for easy switching

5:10 - Quick tip: /model command

/model sonnet - Switch to mid-tier
/model opus - Switch to default
/models - See available providers

6:14 - Cost calculator demo

Light user: $200 → $70 (65% savings)
Power user: $943 → $347 (63% savings)
Heavy user: $2,935 → $1,000 (66% savings)
Link: https://calculator.vlvt.sh

8:04 - Why I don’t use free tiers

Aggressive rate limits
Slow performance (congested)
Can disappear without notice
Better: “almost free” paid tiers ($0.40-0.50/M tokens)

9:17 - Final tips & next steps

📝 Notes & Key Takeaways

Main Insights

Default OpenClaw is expensive: Everything routes to your primary model (usually Opus), including heartbeats, sub-agents, and simple lookups.
Model tiering is the solution: Different tasks need different intelligence levels:
- Heartbeats (periodic checks) → Gemini Flash-Lite ($0.50)
- Sub-agents (parallel work) → DeepSeek R1 ($2.74)
- Complex tasks → Keep Opus ($30)
Massive cost savings: 50-80% reduction with proper configuration, no quality loss on important tasks.
Provider fallback matters: If Anthropic rate-limits, falling back to Sonnet won’t help. Fallback to a different provider (GPT-5.2, Gemini) keeps you running.
Cheap models are fast: Gemini 3 Flash runs at ~250 tokens/sec vs Opus at ~50 tokens/sec.
Free tiers are risky for production: Rate limits, slow performance, can disappear. “Almost free” paid tiers ($0.40-0.50) are worth the reliability.

Actionable Points

Update your config file (~/.openclaw/openclaw.json):
- Set heartbeat model to Gemini 2.5 Flash-Lite
- Set sub-agent model to DeepSeek R1
- Configure fallback chain with different providers
- Add model aliases for easy switching
Use the cost calculator (https://calculator.vlvt.sh):
- Input your usage pattern (heartbeats, sub-agents, queries)
- See exact savings potential
- Get copy-paste config output
Master the /model command:
- Switch models on-the-fly for cost control
- Use aliases (opus, sonnet, flash) instead of full paths
- Stay on Opus for complex work, switch to cheaper for quick tasks
Restart gateway after config changes:
```
openclaw gateway restart
```
Monitor your usage: Track which tasks use which models to optimize further.

Personal Reflections

Add your own notes here after watching

⭐ Rating & Review

After completion:

Quality (1-5): 5/5
Relevance (1-5): 5/5
Would recommend: Yes
Best for: OpenClaw users with API cost concerns, power users running multiple agents

🏷️ Auto-Generated Tags

Content Analysis:

Type: video (YouTube content)
Topics: AI agents, OpenClaw configuration, cost optimization, model selection
Complexity: Technical but accessible - requires basic understanding of OpenClaw and API concepts
Priority: High - immediate actionable cost savings

Why These Tags:

AI - Core subject matter
openclaw - Specific platform discussed
productivity - Efficiency improvement focus
cost-optimization - Main theme of the video
technical - Requires configuration and technical setup
actionable - Clear, implementable steps provided
tutorial - Step-by-step guidance format
inbox - New content requiring action

Suggested Bases Filters:

Find similar content: type = video AND tags contains "cost-optimization"
Find high-priority learning: priority = high AND status = inbox
Find OpenClaw tutorials: tags contains "openclaw" AND tags contains "tutorial"

Related Searches:

OpenClaw documentation on model configuration
DeepSeek R1 vs Claude Opus benchmarks
OpenRouter auto-routing configuration
API cost monitoring tools for AI agents
Multi-agent system architecture patterns

Useful Links:

Cost calculator: https://calculator.vlvt.sh
Config file: https://velvetshark.com/openclaw-multi-model-routing
OpenClaw docs: https://docs.openclaw.ai/
OpenRouter: https://openrouter.ai

Model Pricing Reference (Feb 2026):

Gemini 2.5 Flash-Lite: $0.50/M tokens
DeepSeek V3.2: $0.53/M tokens
GLM 4.7: ~$0.40/M tokens
DeepSeek R1: $2.74/M tokens
GPT-5: $11.25/M tokens
Claude Opus 4.5: $30.00/M tokens

Captured: 2026-02-06 Source: https://www.youtube.com/watch?v=fkT41ooKBuY Channel: VelvetShark

Connection to Other Notes: This video directly relates to OpenClaw operational efficiency and cost management. Consider cross-referencing with:

Your OpenClaw configuration documentation
API cost tracking spreadsheets
Other VelvetShark tutorials on OpenClaw optimization