How to reduce your AI agent running costs with smart model routing

The cost optimization problem

A personal AI agent running on API-based models has a variable cost that directly tracks usage. For light users, this is barely noticeable — a few dollars per month. For heavy users who use their agent for research, drafting, analysis, and automation throughout the day, costs can reach EUR 20–50/month or more if they're using frontier models (GPT-4o, Claude 3.5 Sonnet) for everything.

The good news: most interactions don't require frontier model quality. A model that costs 20x less often produces equivalent results for everyday tasks. The key is routing intelligently rather than using a single model for everything.

The task-quality matrix

Different tasks genuinely require different capability levels. Before optimizing, it helps to categorize your typical requests:

Task Type	Requires Frontier?	Recommended Model Tier
Quick factual Q&A	No	Fast/cheap (GPT-4o mini, Haiku)
Email drafting (simple)	No	Fast/cheap
URL summarization	No	Fast/cheap or DeepSeek
Complex code review	Usually	Frontier (Claude Sonnet, GPT-4o)
Strategic analysis	Usually	Frontier
Creative writing	Depends on quality bar	Either, based on preference
Data extraction/parsing	No	Fast/cheap
Morning briefing generation	No	Fast/cheap

If 60% of your interactions fall into the "fast/cheap" category, routing them to GPT-4o mini instead of GPT-4o reduces that portion of your bill by roughly 25x.

OpenClaw's model routing configuration

OpenClaw supports a tiered model configuration where you define:

• Default model: Used for all requests unless overridden (set this to a cheap model)
• Power model: Used when you explicitly ask for deeper analysis
• Fast model: Used for scheduled automations (morning briefings, cron tasks)

Example configuration in openclaw.config.json:

{
  "models": {
    "default": "openrouter/deepseek/deepseek-chat",
    "power": "openrouter/anthropic/claude-3-5-sonnet",
    "fast": "openrouter/openai/gpt-4o-mini"
  }
}

Then in Telegram, you can invoke the power model explicitly when you need it:

!power Review this contract clause and identify any unfavorable terms...

Everything else routes through DeepSeek V3, which handles most tasks admirably at a fraction of the cost.

DeepSeek V3: the cost-performance sweet spot

DeepSeek V3 (available via OpenRouter) has become the default recommendation for personal agent usage. At USD 0.27/1M input tokens and USD 1.10/1M output tokens, it costs approximately 10x less than GPT-4o while producing results that are competitive or superior on most everyday tasks.

Where DeepSeek performs well: summarization, drafting, Q&A, code generation, data analysis, translation.

Where frontier models still have an edge: nuanced creative writing, highly complex multi-step reasoning, tasks requiring very recent knowledge.

Prompt compression

Every token in your prompt costs money. Long system prompts, verbose instructions, and unnecessary context all add up. A few practices that reduce token consumption without degrading quality:

• Concise system prompt: Review your system prompt for redundancy. "You are a helpful assistant who is always polite and professional and responds with kindness and care" can become "You are a professional assistant. Be direct and concise."
• Context trimming: Configure OpenClaw to include only the last N messages in each API call for routine interactions, rather than the full conversation history.
• Summarize don't repeat: When sharing long documents, paste only the relevant excerpt rather than the full text.

Caching for repeated queries

If you have scheduled automations that make similar requests repeatedly (same system prompt, same structure), OpenAI and Anthropic both offer prompt caching that discounts tokens used in the cached portion. OpenClaw's scheduled tasks are designed to take advantage of this automatically when the API supports it.

For Anthropic's APIs, caching prefix tokens that are identical across calls can reduce costs by up to 90% for the cached portion. This makes scheduled morning briefings and regular summaries significantly cheaper over time.

Monitoring your spend

Set up spend alerts on your OpenRouter or API provider dashboard. Most providers let you set a monthly spend limit with email notification at a threshold (e.g., alert at USD 5, hard cap at USD 20). This prevents surprise bills from runaway automations or accidental loops.

Your k-claw dashboard shows API call volume by day and estimated cost trends, so you can identify unexpectedly expensive workflows before they accumulate.

Local models for zero marginal cost

The ultimate cost optimization is routing simple queries to a local model via Ollama. If you're already paying for a larger VPS (8+ GB RAM), running Llama 3.1 8B or Mistral 7B locally means that category of requests costs EUR 0 in API fees.

A practical hybrid approach: use a local 7B model for quick Q&A and routine tasks, OpenRouter DeepSeek for medium complexity, and Claude/GPT-4o for your most demanding work. This three-tier setup can bring monthly API costs below EUR 5 for most users while maintaining excellent quality where it matters.