How to compare reasoning vs fast models

Pick the right model tier with a tiny eval — not hype.

1. Collect 10 hard + 10 easy real tasks from your product

2. Run both model tiers with the same prompt template

3. Score: correct, helpful, latency, cost

4. Promote reasoning only where it wins clearly on hard set

5. Keep fast model as default for volume paths

**Try the lesson:** `reasoning-models` in Lane C