API pricing is quoted per million tokens — separately for input, for output, and (increasingly) for reads and writes of the prompt cache. The multipliers are modest-looking until you multiply by volume.

This calculator counts tokens with the cl100k_base tokenizer (a good stand-in for most modern models) and runs the math across a small curated set of models. Everything happens in your browser. Prices are frozen to the date at the top of the breakdown — always verify against each provider's current pricing page before committing a budget to it.

What moves the bill

Output is the expensive part

At Claude Opus 4.7's rates, an output token costs five times more than an input token. Type a long prompt, set output to 100 tokens, and the input still dominates. Flip output to 5000 tokens — output overtakes it immediately. If you're optimizing cost, shorter answers save more than shorter prompts.

Cache hits move the needle

With a 50% cache hit rate on a long prompt, Anthropic's cached-read price is one-tenth the input rate. On high-volume pipelines with stable system prompts, caching can cut the monthly bill by 40–60%. Slide the cache hit rate up and watch the per-call cost drop. Then remember that caching needs you to keep the prompt prefix byte-for-byte stable across calls.

Model tier > clever engineering

At identical input+output, Haiku 4.5 is roughly 15× cheaper than Opus 4.7. The hardest cost-optimization question is usually "am I on the right tier?" — not "can I shave 10% off the prompt?"

Caveats

The tokenizer estimates input tokens with cl100k_base. Anthropic's tokenizer is close but not identical — assume ±10% error on the token count. Output token counts are whatever you specify. Cache prices model the common provider pattern (a cheap read rate + a writeable premium); specific cache semantics differ per provider, so the number here is a planning estimate, not an SLA.