AutoRouter

haiku's tokens are 40× cheaper than opus but its tasks only 25× — it takes more turns to get there on the refactor: 8.7 turns vs opus's 4.3. wall time stayed within 9 seconds across all three models though

most of that opus cost isn't the model writing — it's the setup 69% goes to writing the system prompt, tool definitions & files into the cache before a single token comes back a fresh session pays that fixed cost every time

haiku 5.5 runs a claude code task for $0.0044 opus runs an average task for $0.110 — 25× more all 45 runs passed. on these small tasks, the cheap model passed every check

How AutoRouter picks a model for each turn, in order: 1. Your rules 2. How hard the turn is 3. Only models that clear the quality bar for it 4. Your ChatGPT & Claude plans first for hard work 5. The cheapest capable model for everything else Every turn shows the model, what it cost and why. It sometimes misjudges difficulty. Rate the pick and we retune.