Estimate your monthly cost for Anthropic's Claude models — Fable, Opus, Sonnet and Haiku.
| Model | $/1M input | $/1M output | Context |
|---|---|---|---|
| Claude Fable 5 | 10.00 | 50.00 | 1M |
| Claude Opus 5 | 5.00 | 25.00 | 1M |
| Claude Sonnet 5 | 2.00 | 10.00 | 1M |
| Claude Haiku 4.5 | 1.00 | 5.00 | 200K |
Output tokens cost 5× input, so trimming verbose responses saves the most. Prompt caching can cut the cost of repeated context (system prompts, long documents) to a fraction of the standard input rate. And route easy tasks to Haiku — it's 10× cheaper than Opus and often good enough.
Most bills are dominated by input, not output. A production request rarely sends a short question: it carries a system prompt, tool definitions, retrieved documents and conversation history, all of which are resent on every call. The visible part — what the user typed — is usually the smallest component.
That has a practical consequence. Shortening responses feels like the obvious saving because output is priced higher per token, but if input outweighs output by ten to one, trimming context saves more money than trimming answers.
Prompt caching. Where the same prefix — system prompt, tool schemas, a long document — is sent repeatedly, cached reads are charged at a fraction of the standard input rate. For agents and chat products that resend the same preamble on every turn, this is usually the single largest reduction available, and it needs no change to what the model actually does.
Model routing. The gap between tiers is large enough that sending every request to the most capable model is rarely justified. Classification, extraction, routing and summarising are handled well by the smaller tiers; reserve the top tier for work that genuinely needs it. Measure quality on your own task rather than assuming.
Batch processing. Work that does not need an answer in the next few seconds — backfills, evaluations, bulk enrichment — can be submitted asynchronously at a reduced rate. The trade is latency, which for offline jobs costs nothing.
Cost per request is input tokens times the input rate, plus output tokens times the output rate, divided by a million. Multiply by expected monthly volume. The number that most often turns out wrong is not the rate but the token count: measure a real request with the Token Counter rather than estimating from word count, and check that your context fits comfortably using the Context Window Calculator.
Claude Haiku 4.5, at $1/$5 per million tokens. Claude Fable 5 is the priciest listed here at $10/$50.
Use the Token Counter, then plug the numbers in above.
No — you are billed for the tokens you send, not for the size of the window available. A model advertising a million-token window costs the same as a smaller one at equal usage. The risk is behavioural: a large window invites sending more context than the task needs, and that does cost more.
The usual causes are retries after failures, conversation history growing across turns, and system prompts or tool definitions that were never counted. Long-running chat sessions are the classic surprise: the same conversation costs progressively more per message as the history it carries gets longer.
Compare across providers: LLM API Cost Calculator · GPT · Gemini