LLM cost calculator for AI agents

Multiply the tokens in each model call by the model’s price per token, then by the number of calls (steps) in a run. Input, output and cached input tokens each have their own price, so they are priced separately: cost per step = input × input price + cached input × cache price + output × output price. Cost per month is cost per run × runs per day × days.

From LiteLLM’s open model price list, which the LiteLLM gateway uses to bill requests and which its maintainers and community update as providers change prices. This page re-reads it every day, so a price change shows up without anyone editing the page.

An agent makes many model calls per task and resends the whole conversation on each one: the system prompt, tool definitions, every earlier tool result. Input therefore grows with every step. In Detailed mode, “context growth per step” models this, and the peak context shows how close a run gets to the model’s context window.

Simple mode takes one tokens-per-step number and splits it the way a typical agent step looks: 90% input, 10% output, and 70% of input served from the prompt cache when the model has a cache price. Detailed mode lets you set input, output, context growth, cache hit rate and days per month yourself. Switching to Detailed carries your Simple numbers over.

Providers charge much less for input they have already cached, often 10% of the normal input price. Agents resend a stable prefix every step, so cache hit rates of 60–90% are common. Some providers, such as Anthropic, also charge a premium to write tokens into the cache; the calculator bills uncached input at that write price when caching is on.

Reasoning tokens are billed as output tokens, so include them in output tokens per step. For reasoning models they are often the largest part of the output.

Some models, such as Gemini Pro and recent GPT models, charge a higher rate once a single request goes over a size threshold (for example 200K input tokens). The calculator applies that rate to each step that crosses the threshold and tells you how many steps did.

No. The calculator uses list prices for standard requests. Batch APIs usually cost about half, and negotiated enterprise rates vary.