Start with the exact model name
Anthropic's pricing page lists individual Claude models and billing categories. Record the model ID returned by your API and the date of the price sheet. A family nickname like “Sonnet” is not enough to quote a reliable rate because versions can differ.
Estimate a standard request
Get input and output token counts from the Anthropic usage record. Multiply each by its current per-million-token rate and add. For illustration, 2,000 input at $3 per million and 500 output at $15 per million totals $0.0135. Those sample rates are arithmetic examples; confirm the applicable live model rate before budgeting.
Account for cache and batch
Anthropic separates cache creation and cache read billing from ordinary input. Repeated context may save money only when the cache is actually hit. Its batch feature can change eligible processing rates. Add these categories from the usage report rather than charging all input at one price.
Compare models on your tasks
Test a cheap, mid-range and high-capability Claude option on the same representative requests. Track pass rate, average output length, latency and retries. A model with a higher per-token rate may still be economical when it reduces rework. Use the calculator to price the measured counts.