What it looks like
Finding out why a build fails — 30,000 input tokens (code and log):
| Option | Attempts to a correct answer | Tokens in total | Cost | Your time |
|---|---|---|---|---|
| Haiku 5.5 | 4 plus a hint | ~140K in, 8K out | ~$0.02 | 20 minutes checking wrong answers |
Sonnet 5.5, high |
1–2 | ~60K in, 10K out | ~$0.22 | 5 minutes |
Opus 5.5, low |
1 | ~30K in, 3K out | ~$0.18 | 2 minutes |
The attempt counts are an example, not a measurement: every task has its own. The point is that tokens here cost cents, while checking a wrong answer costs minutes of your work.
How to compare on your own tasks
- Take 5–10 typical tasks with a known correct result.
- Run them on two or three options: the small model, the mid-tier, the flagship on
low. - Use
usage(Tokens) to count the cost up to a correct answer, and note where you had to ask again. - For each kind of task, keep the cheapest option that solves it on the first try.
Where cheap does pay off
Thousands of uniform requests with a checkable result: classification, field extraction, translating short strings. There a mistake is caught automatically, and a 20–40× price difference multiplies by the volume. That’s also where batching (Non-urgent API requests go as a batch at half price) and caching (Prompt cache) come in.