AI agent cost calculator

How the calculator estimates your bill

An AI agent is a program that relies on a language model to complete a task in several steps: reading an email, looking something up, drafting a reply, calling a tool. At each step it sends text to the model and gets text back. Providers bill that traffic in tokens, small chunks of words (the token estimator explains the idea in more detail).

The result above comes from a simple formula:

monthly cost = tasks per day × days per month × (input tokens × input price + output tokens × output price) / 1,000,000

Prices are published per million tokens, hence the division by 1,000,000. The cached share of calls then lowers the cost of the input tokens involved: when your requests start with the same text (fixed instructions, a reference document), the provider may bill that part at a reduced rate, if the model supports it.

A worked example with variables

Picture an agent that handles T tasks a day, D days a month. Each task sends I tokens and receives O tokens. With an input price Pi and an output price Po per million tokens, the cost is T × D × (I × Pi + O × Po) / 1,000,000. Double T and the bill doubles. Halve O and only the output share drops. The tool applies exactly this logic, model by model, using public prices.

Why output usually costs more than input

With most providers, a generated token (output) costs more than a token the model reads (input). The reason is technical: the model processes the input in one pass, but it writes its answer one token at a time, and each token needs a fresh round of computation. Writing simply uses more resources than reading. In practice, an agent that drafts long reports can cost more than one that reads a lot but answers in a few lines.

How to estimate tokens per task

The presets (email assistant, customer support agent, research agent, coding agent) give a realistic starting point for each profile. To refine them:

• Input: add up the system prompt (the fixed text describing the agent’s role), the user’s request, any documents or tool results injected, and the conversation history. A multi-step agent often resends the whole context on every call, so count it every time.
• Output: estimate the usual length of the answer, plus intermediate text such as plans and tool calls.
• Reasoning: so-called reasoning models sometimes produce internal thinking that is billed as output. Leave some headroom.

The most reliable approach is to measure a real exchange with the token estimator and plug the numbers in here.

Reading the results

The table ranks models from cheapest to most expensive for your scenario. The capability filters exclude those that cannot do what your agent needs: tools (calling external functions), vision (reading images) or reasoning (thinking step by step before answering). The chart shows the 15 cheapest compatible models. The “best score-to-price” indicator weighs cost against benchmark results, the standardized tests explained in the AI ranking. Keep in mind that a very cheap but unreliable model can end up costing more if tasks need to be rerun or fixed.

How to cut the bill

• Use caching: put fixed instructions and recurring documents at the start of the prompt, always in the same order. The higher the cached share, the cheaper your input.
• Pick a smaller model for simple jobs: sorting emails or pulling out a date does not need the most powerful model. Many agents pair a small model for triage with a larger one for hard cases. The model comparison tool helps you choose.
• Keep the context lean: summarize the history instead of resending everything, and attach only the relevant parts of a document.
• Constrain the output: ask for short, structured answers whenever you can.
• Cap the steps: an agent that loops without a limit multiplies calls, and the bill with them.

Limits of the estimate

The prices used are public list prices from models.dev, cross-checked against OpenRouter, excluding tax. They can change at any time and do not include negotiated discounts, limited free tiers, add-on tool fees (web search, storage) or retries after errors. The data is refreshed every week, and the methodology page lists the sources. Treat the result as a ballpark for comparing options, not as a price guarantee.