Claude API prices per million tokens
| Model | Input | Output | Cache write | Cache read |
|---|---|---|---|---|
| Opus 5.5 | $4 | $20 | $5 | $0.2 |
| Sonnet 5.5 | $2 | $10 | $2.5 | $0.1 |
| Fable 5.1 | $10 | $50 | $12.5 | $0.25 |
| Haiku 5.5 | $0.1 | $0.5 | $0.125 | $0.01 |
Source: claude.com/pricing, October 2026. Five-minute caching; Haiku 5.5 is higher above 100K-token prompts.
Why Claude Code bills look the way they do
An agent re-reads the conversation on every step. Claude caches that history, so most tokens in a long session are cache reads, which are cheap one by one but add up. Output is the most expensive token, and the one you control least.
| A day of Opus 5.5 | Tokens | Cost |
|---|---|---|
| Input | 0.5M | $2.00 |
| Output | 0.4M | $8.00 |
| Cache writes | 3M | $15.00 |
| Cache reads | 40M | $8.00 |
Plan or API?
If the same usage costs more on the API than a plan, and it fits in the plan's limits, the plan wins. The catch is the limits: a heavy day can use a five-hour window early. Timecard reads your real sessions on your Mac and shows the API value of each one, so you know which side of the line you're on.