The Real Cost of an AI Coding Agent: A Token Budget Breakdown

Why an agent's bill doesn't look like a chat bill

A single chat turn sends a prompt and gets an answer. An agent turn sends the accumulated conversation, the tool schema, a slice of the repository, the output of the last tool call, and often a self-critique pass before it commits to an action -- and it does this on every step of a loop that can run for dozens of steps to complete one task. Spheron's analysis of agentic inference puts the resulting multiplier at 5-30x the token volume of an equivalent chat interaction, a range wide enough to reflect how differently one agent framework's loop is built from another's, but consistently far above one-shot chat regardless of where in that range a given framework lands [2].

That multiplier is the reason a per-token sticker price is a poor predictor of what an agent actually costs to run. Two models priced identically per token can produce very different monthly bills if one framework's loop resends more context per step, retries more often, or runs more steps to reach the same result -- none of which shows up on a rate card at all.

Where the tokens actually go: review, not writing

SaaSCity's analysis of the ChatDev multi-agent framework breaks that loop down by stage, and the finding is specific: 59.4% of total token consumption in the framework studied went to code review -- re-reading the diff, checking it against the original request, running a critique pass -- rather than to generating the code itself [3]. That is one framework's measurement, not a law every coding agent obeys, but it is directionally consistent with why the tokens an agent burns skew so heavily toward re-reading context it has already seen rather than producing new text.

That skew is exactly the shape prompt caching is built for. A review pass that re-reads the same system prompt, the same tool schema and most of the same repository slice as the step before it is, by definition, resending a prefix -- and a resent prefix is what a cached rate exists to discount.

What a closed-frontier day costs, and why the comparison matters

Morph's token-math comparison of coding-agent spend states that a full day's usage pattern on a closed frontier model runs roughly $594 a month at the rates it cites -- a comparison point attributed to that piece, not a price this platform charges or a model this platform sells [1]. What makes it relevant here is the mechanism behind it: a bill that size is not explained by the per-token rate alone. It is explained by a review-heavy loop resending large amounts of context, at a rate card that, on a closed frontier model, typically has no cached-prefix discount anywhere near as deep as an open-weight rate table's.

That is the specific reason cached-prefix pricing matters more for an agent than the flat sticker price per token does: the review overhead SaaSCity measured and the token multiplier Spheron measured both describe a workload where most of the bill is repeated context, and repeated context is the one part of a bill a cache discount actually reaches.

A heavy-agent month, priced against the official rate

Turning that into a number means stating an assumption rather than measuring it: 480 million input tokens and 12 million output tokens a month, weighted roughly 40 to 1 toward input -- the shape a review-heavy agent loop produces once most of its volume is repeated context and critique passes rather than newly written code, consistent with the review-overhead and token-multiplier findings above rather than derived from them directly.