LLM price calculator: cost per task, per day, per month

How to use it

1. Estimate the size of a task. Paste a representative prompt (instructions, context, document) into the text box: the tool counts characters and words and derives an approximate token count. Do the same with a typical answer for output tokens.
2. Enter the volume. The number of tasks per day: assistant messages, tickets handled, files summarised, agent steps. A month counts 30 days.
3. Add caching if you use it. If a fixed part of the prompt (system instructions, documentation) is reused on every call, enter its share: it is billed at the cache rate when the model publishes one.
4. Set a minimum level. The ECI filter removes models that are too weak: there is no point comparing a small model's price with a frontier model's if the first cannot do the task.

The result updates as you type: cost per task, per day and per month, cheapest first, with the monthly token volume. All calculation runs in your browser; pasted text never leaves the page.

How the calculation works

Providers bill separately for tokens sent to the model (input) and tokens it produces (output), in dollars per million tokens. Output usually costs four to eight times more than input. The cost of one task is therefore:

cost = (uncached input × input price + cached input × cache price + output × output price) ÷ 1,000,000

The daily cost multiplies this amount by the number of tasks, the monthly cost by 30. For models with no published cache price, all input is counted at full price.

The token counter uses the rules of thumb published by OpenAI for English: one token is about 4 characters, and 100 tokens about 75 words. The tool keeps the larger of the two estimates, because languages with longer words and accents produce slightly more tokens than English. Source code and non-Latin scripts produce noticeably more.

The case of AI agents

An agent does not make one call per task but runs a loop: it reads, calls a tool, reads the result, starts again. At every turn, the whole history is sent back to the model. A twenty-step coding-agent task with a context growing to 60,000 tokens therefore consumes several hundred thousand input tokens. To estimate an agent, count the average number of steps and multiply the average context size by it, or better: read the real usage of one session in your provider's dashboard and enter it here.

This is where prompt caching changes everything: the start of the history is identical from one turn to the next and can be billed at the cache rate. 232 of the 356 models listed publish such a rate. The guide Reduce costs covers the other levers.

Where the prices come from

The 356 prices come from OpenRouter's public API, read on October 25, 2026. For each model, it is the reference provider's price, excluding tax. Free variants (8 in the list) are hidden by default: they are rate-limited and unsuitable for regular use. The ECI shown next to the price comes from Epoch AI; it helps you avoid comparing models of very different levels.

What the calculation does not tell you

• Reasoning tokens. Models that “think” before answering bill those hidden tokens as output. Depending on the effort requested, real output can be several times longer than the visible answer.
• Tokenizers differ. The same text does not give the same token count from one model to another; the gap can reach 10 to 30%.
• Discounts. Batch processing, volume commitments and agent subscription plans (a monthly fee instead of per-token billing) are not taken into account.
• Tools billed separately. Web search, code execution or image inputs sometimes have their own price.

Treat the result as an order of magnitude, then check it against a week of real usage. To go further: LLM API pricing.