Five rate categories, not one token price
The following selected model rates come from the Claude API pricing documentation, checked on September 30, 2026. Amounts are USD per million tokens on the native Claude platform. Other provider contracts, taxes, and separately charged tools can change the invoice.
| Model | Input | 5-min cache write | 1-hour cache write | Cache hit | Output |
|---|---|---|---|---|---|
| Sonnet 5.5 | $2.00 | $2.50 | $4.00 | $0.10 | $10.00 |
| Opus 5.5 | $4.00 | $5.00 | $8.00 | $0.20 | $20.00 |
| Fable 5.1 | $10.00 | $12.50 | $20.00 | $0.25 | $50.00 |
| Haiku 4.5 | $1.00 | $1.25 | $2.00 | $0.10 | $5.00 |
Cache creation is not priced like a cache hit. A longer retention choice costs more to write, so it needs sufficient reuse to make sense for the workload. Keep the categories distinct when using reported consumption; do not charge the same input category twice.
The same token assumptions, four costs
Hypothetical example: 100,000 uncached input tokens and 10,000 billable output tokens, normal processing, no cache creation, no batch discount, and no extra tool charges:
- Sonnet 5.5: 0.1 × $2 + 0.01 × $10 = $0.30.
- Opus 5.5: 0.1 × $4 + 0.01 × $20 = $0.60.
- Fable 5.1: 0.1 × $10 + 0.01 × $50 = $1.50.
- Haiku 4.5: 0.1 × $1 + 0.01 × $5 = $0.15.
This is an illustrative calculation, not an observed bill and not an equal-quality benchmark. Actual token counts, retries, reasoning, and accepted results can differ. Text length alone is not a substitute for the usage returned by the service.
Batch, fast processing, and long context
The pricing documentation lists a 50% Batch discount on input and output. Opus 5.5 fast processing has higher $8 input and $40 output rates and does not combine with Batch. Sonnet 5.5 and Opus 5.5 have standard pricing across their full one-million-token windows; do not transfer an older model's long-context surcharge to these models.
Nor should you assume every model in the table has the same context limit. A table comparing prices is not a capability matrix. Before choosing a model for a large input, check its own model reference and the conditions of your deployment.
Price parity is not task parity
The GPT price snapshot lists GPT-6.1 Sol at the same base input, output, and cached-input rates as Sonnet 5.5. These matching rates do not tell us which will use fewer tokens or produce fewer revisions for a given job. For the more expensive Opus tier, see our direct GPT-6.1 Sol versus Opus 5.5 comparison.
For a fair comparison, preserve the source material and acceptance criteria, then measure the total spend to reach an accepted result. Include the rejected attempts. If a lower-priced model needs multiple repairs, the nominal input price has answered only part of the budgeting question.
Subscriptions are a separate purchase
The Pro plan documentation explicitly excludes Console API usage. A Claude subscription therefore should not be treated as prepaid access to the API rates above. Review the Pro/Max usage article for included allowances and model-specific exceptions.
A useful monthly forecast starts with measured API usage, planned accepted tasks, and an allowance for retries. Recheck the official page before renewing a budget: this article is a dated snapshot, not a permanently fixed price list.