A small model beats a frontier model whenever the answer can be checked by code, because the check costs less than the upgrade.
The price gap is enormous and the capability gap is not. Haiku 5.5 costs $0.10 per million input tokens; Opus 5.5 costs $4. That is forty times. On the Artificial Analysis Intelligence Index, the top three models sit at 59, 58 and 57 — within two points of each other, across three labs.
The check is the whole trick
If a function returns a schema, a compiler accepts the code, or a test suite goes green, a wrong answer is caught for free. Then you run the cheap model, and you run it again when it fails.
If a person has to read the output to know whether it is right, you have moved the cost, not removed it.
Four cases where small wins
- Closed label space: classification, routing, tagging. There is nothing to reason about.
- Extraction into a fixed schema, with validation that rejects bad rows.
- High volume at low stakes: the price difference becomes the budget.
- Latency-sensitive paths, where a fast small model beats a slow good one.
Two cases where it loses
Long agent runs, where one bad step poisons the next ten and a retry means redoing the work. And open-ended writing, where nobody can say what "correct" is, so nothing catches the miss.
| Task | Cheap model | Why |
|---|---|---|
| Ticket routing | Yes | Fixed labels, cheap to spot-check |
| JSON extraction | Yes, with validation | The validator is the test |
| 20-step code migration | No | One wrong edit costs the run |
| Customer-facing prose | No | No automatic check exists |
How to find out
Run both on 50 real tasks from your own logs, and grade with the check you already have. If the cheap model passes, keep it. If it fails 3 of 50, the frontier model is not a luxury; it is the cheaper option.