---
name: model-cost-optimization
description: Best practices for reducing expensive model token usage
author: Hermes Agent
tags: [cost-optimization, model-selection, efficiency]
---

# Model Cost Optimization Guide

## When to Use Cheap vs Expensive Models

### Use Cheap Model For:
- Simple factual questions and basic information
- Short clarifications (under current thresholds)
- Basic tool parameter preparation
- Simple yes/no decisions
- Routing questions like "which skill should I use?"

### Use Expensive Model For:
- Complex multi-step reasoning
- Code generation and debugging
- Long-form content creation
- Advanced planning and analysis
- Creative tasks requiring originality

## User Query Optimization Tips

1. **Break down complex requests** into smaller, focused questions
2. **Be specific and direct** in your queries
3. **Ask one thing at a time** rather than multiple questions
4. **Use clear, concise language** to stay under character limits

## Current System Behavior

The system automatically routes based on:
- Message character count
- Word count  
- Current thresholds are set to maximize cheap model usage while maintaining quality

If you find responses from the cheap model are insufficient, you can:
- Rephrase your query to be more complex (triggering expensive model)
- Explicitly request detailed/complex analysis
- Ask follow-up questions that naturally exceed thresholds

## Hermes Delegation Model Strategy

When using `delegate_task` for subagents, configure a cheaper/faster model to save cost without sacrificing quality on simple tasks.

### Setup

```bash
hermes config set delegation.provider deepseek
hermes config set delegation.model deepseek-v4-flash
```

API key auto-resolves from `.env` (e.g. `DEEPSEEK_API_KEY`).

### Decision Rules — Delegate vs Handle Yourself

| Delegate to cheap model | Keep in main (expensive) model |
|---|---|
| Translation, formatting, boilerplate | Complex architecture design |
| Simple code snippets (≤1 function) | Multi-step debugging |
| File batch operations (search, rename, extract) | Tasks needing deep context understanding |
| Data extraction and parsing | Tasks requiring iterative back-and-forth |
| Mechanical, well-scoped tasks | Ambiguous tasks with unclear requirements |

**Litmus test:** Can you describe the task in one sentence and does it need no iterative feedback? → delegate to cheap model.

### Pitfalls

- **Subagent self-reports are unreliable** — verify file creation, HTTP responses, and test results yourself after the subagent finishes
- **Don't delegate tasks needing user clarification** — subagents can't use `clarify`
- **Flash models are more verbose** — they may use more tokens per task but are still cheaper overall due to lower per-token pricing
- **Don't delegate when context is critical** — subagents have no memory of your conversation, so pass all relevant file paths, error messages, and constraints explicitly

## Monitoring Effectiveness
See `references/deepseek-v4-benchmark.md` for real-world speed/token benchmarks of DeepSeek models.

Pay attention to response quality. If simple tasks are getting inadequate responses, the thresholds may need adjustment through configuration changes.