Free tool
LLM cost calculator
Estimate your monthly LLM bill and compare it across models. Enter your per-request token usage and volume — everything runs in your browser.
Same workload, every model
Sorted cheapest first. A price difference is not a saving until the cheaper model passes your quality bar on these prompts.
| Model | Monthly cost | vs. selected |
|---|
Anthropic figures verified September 25, 2026; others June 2026. Prices change frequently. Always confirm against the provider’s official pricing page before relying on these figures for budgeting.
Frequently asked questions
How is LLM API cost calculated?
LLM cost is charged per token: (input tokens × input price) + (output tokens × output price), priced per million tokens. Multiply by your monthly request volume to estimate the monthly bill. This calculator does that across models so you can compare.
Why is output so much more expensive than input?
Providers price output tokens 3–5x higher than input because generation is more compute-intensive than reading the prompt. That makes output length — and reasoning tokens, which bill as output — a major cost driver worth optimizing.
How do I actually reduce these costs?
Right-size the model per route, cache stable prompt prefixes, batch non-urgent work, and add fallback chains — all behind evals so quality holds. How much that saves depends on your traffic, so measure a baseline first. We do this as a fixed or outcome-priced engagement.
What this estimate includes, and what it leaves out
The arithmetic here is deliberately simple: requests multiplied by tokens multiplied by the per-million price for the model you picked, with input and output priced separately because output typically costs several times more. That is enough to size an idea, sanity-check a vendor quote, or work out whether a feature is viable before anyone builds it. It is not a forecast, and the gap between this figure and a real invoice is usually explained by the same handful of things every time.
Retries and fallbacks are the first. A request that times out and is retried bills twice, and a fallback to a larger model on failure bills at the larger model's rate — so a route with a 5% failure rate is not 5% more expensive, it is more than that once the fallback is priced in.Prompt caching pulls the other way and is not modelled here: where a long, stable prefix is reused across calls, cached input reads are billed at a fraction of the standard rate, which can move a high-volume route substantially. Reasoning tokens bill as output on models that produce them, and they are invisible in any estimate based on the length of the answer the user sees.
The largest omission is agent loops. This calculator assumes one request equals one model call, which holds for a classification or a summarisation and breaks completely for anything agentic. An agent re-sends its accumulated context on every step, so a task that takes fifteen steps costs far more than fifteen single calls — the context grows as it goes, and cost grows with it. If your workload is agentic, estimate per completed task by measuring a handful of real runs rather than per call.
Token counts are also approximate by nature. Tokenisers differ between providers, so identical text is a different number of tokens depending on who you send it to, and code, non-English text, and structured output all split less efficiently than ordinary prose. For anything you intend to budget against, count with the target model's own tokeniser on a real sample of your traffic. And treat the prices themselves as a snapshot: they are checked against provider documentation on the date shown and change without notice, so confirm against the official pricing page before a figure from this page reaches a spreadsheet anyone commits to.