AI apps · Usage only · unit: thousand-token bundles
Anthropic Claude API pricing teardown: usage-based rates
Short answer
The Claude API prices per token by model, with separate input, output, cache and batch rates and distinct rates for some prompt-length bands.
Pricing mechanics
| Plan or rate | Position in ladder |
|---|---|
| Standard model pricing | Entry |
| Batch | Paid step |
| Prompt caching rates | Top published step |
- What the docs describe
- Rates vary by model, with distinctions for input, output, cache and batch. Some prompt-length bands have distinct rates.
- Free or trial fence
- Not stated in the linked pricing documentation.
- Overage or upgrade rule
- Pure usage at model-specific rates.
- Cap or spend control (none stated)
- Not stated in the linked pricing documentation.
A feature or configuration fence (what a plan includes) is not the same as an abuse cap. We do not infer private costs, margins or unpublished limits.
Takeaway for founders
Model cache hit rates before quoting one effective cost per action; a long-prompt band can change the rate entirely.
Load into simulator
Illustrative remix · amounts and costs are editable.
Fictional tiers inspired by the pattern above. It does not reproduce Anthropic Claude API's plans, invoices or margins. Fictional variable cost: $0.0035 per thousand-token bundle. A thousand-token bundle is a sample blended unit.
| Tier | Model | Price | Limit | Accounts |
|---|---|---|---|---|
| Sample entry | Usage only | $0.01 per unit | hard cap 20,000 thousand-token bundles | 200 |
| Sample growth | Usage only | $0.008 per unit | hard cap 100,000 thousand-token bundles | 70 |
| Sample scale | Usage only | $0.006 per unit | hard cap 350,000 thousand-token bundles | 15 |
Related guide: Hybrid vs. usage-based billing · All teardowns