How much an AI agent costs a month, counted in tokens

How tokens turn into dollars

A token is a piece of text, a few characters long. Providers sell them by the million and charge four different rates:

• Input: text the model reads for the first time.
• Cache write: the same text, stored so the next call can reuse it.
• Cache read: stored text read again. This is the cheap one.
• Output: text the model writes, thinking included. This is the dear one.

An agent works in a loop: it reads files, runs a command, looks at the result, decides again. Every turn of the loop is one model call, and every call carries the whole conversation so far. That is why input and cache reads, not the answers, make most of the tokens.

The prices, per million tokens

OpenRouter sells the same models through one key: inference is billed at the provider's list price. It adds 5.5% on credit purchases on the Standard plan, 8% on Business, so count about 5.5 percent on top.

What one task costs

We needed a task that fits most coding agents. Ours is a bug fix or a small feature: one coding task: the agent makes 30 model calls. Each call carries about 50,000 tokens of context; 45,000 are read from the cache, 5,000 are new and get written to it, and the model writes 1,000 tokens. That is 1.53 million tokens per task, and the same count for every model below.

Take Claude Sonnet 5.5: 1.35 million cache-read tokens cost $0.135, 150,000 cache-write tokens $0.375, and 30,000 output tokens $0.30. The total is $0.81. Without the cache the same task costs $3.30, about four times more.

What a month costs

Multiply by how much the agent works. Three profiles: light is 20 tasks a month: a side project on weekends; regular is 200 tasks a month: about ten on every working day, which is 306 million tokens; heavy is 1,000 tasks a month: several agents in parallel.

The spread is wide for one reason: a model's price class. The same work costs a few dollars on a small model and hundreds on a top one. Small models fit routine steps. Hard bugs and design work are where the larger ones earn their price, and that is a choice for your own tests, not for this table.

Where the bill grows

• Long sessions. Context grows with every turn, and each call reads all of it again.
• Thinking. Reasoning counts as output, the dearest rate. We put 1,000 output tokens per call; a model that thinks hard writes more.
• Cache misses. A cache entry expires after a short time: five minutes for the cache write priced on Anthropic's page. If the agent waits for you longer, the next call pays for writing the context again.
• Parallel agents. Ten agents in one hour cost ten times one agent, not less.
• Retries. A task that fails and starts over is paid twice.

Ways to pay less

What this page does not say

• The usage is our estimate. Your agent can use a third of this or three times more. Count your own calls for a week before you budget.
• The token count is the same for every model. In practice it differs: Anthropic says Claude 4.7 and later models make about 30 percent more tokens from the same text, so Claude bills may come out higher than the table.
• It covers model tokens only. Search tools, paid APIs your agent calls, the server and taxes are extra.
• Uno AI is not unlimited on Small and Plus: after the hours it stops. Only Max with Uno AI has no end. We do not claim that a given number of hours equals a given number of tokens.
• Prices change. The table is dated 16 October 2026; check the provider pages before you decide.

Questions