AI coding agent pricing: where the money actually goes

The first bill from an AI coding agent is usually a shock, and almost nobody can explain it afterwards. You did what felt like a day's work and it cost more than a month of your editor. The reason is structural, it is the same for every vendor, and once you can see it you can cut your spend by more than any tool switch will.

You are not billed for your question. You are billed for the conversation

A coding agent does not answer once. It reads a file, thinks, runs a command, reads the output, edits something, runs the tests. Each of those is a round trip, and each round trip resends the accumulated conversation so far, because the model has no memory between calls.

This is also why caching exists. Anthropic prices a cache read at a fraction of fresh input precisely because agent traffic resends huge, near-identical prefixes over and over.1 A harness that caches well and one that does not can differ several-fold on the same task with the same model.

Four things that change your bill more than switching vendor

1. Session length. One thread that runs all day is the single most expensive habit. Costs grow with the conversation, so the last turn of a long thread can cost many times the first. Start a new thread when the subject changes - not when the tool complains.
2. How much the harness re-reads. This is a property of the tool, not the model. Some agents re-open the same files every turn. You cannot see this happening and it is most of the bill. It is measurable: watch input tokens per turn.
3. The model. Frontier models cost several times mid-tier ones. Plenty of agent steps - reading a file, running a test, renaming a symbol - do not need the strongest model available.
4. Whether the work was right first time. A run that goes wrong costs twice: once to make and once to undo, plus the turns spent discovering it. Better instructions up front is a cost optimisation, not just a quality one.

Which of the four would you guess is biggest?

For most people it is the first two, and they compound: a long session plus a re-reading harness means every turn resends a conversation that was already bloated. That combination is how a "quick fix" turns into a bill nobody can account for.

The model choice, which is where everyone starts, is usually third.

How to measure it properly, in ten minutes

1. Pick one real task you have done before.
2. Run it on each tool you are considering, from a fresh session.
3. Record total input tokens, total output tokens and whether the result was correct.
4. Ignore wall-clock time. It is not what you are billed for.

Input tokens are the number that matters, because input dominates in agent work and it is where harness design shows up. If one tool sends four times the input for the same result, no model price makes up the difference.

Why the pricing pages are hard to compare

Most vendors are now somewhere between: a subscription with a usage ceiling. That is more honest than it looks, as long as the ceiling is stated in terms you can reason about before you hit it.

Did that land?

Four questions. They test the mechanism, not the numbers.

Why does a 40-step task cost more than 40 single questions?

Each step escalates to a larger model as the task gets harder Every step resends the whole conversation, which keeps growing Steps inside an agent run are billed at a premium tool-use rate Output tokens cost several times more than input tokens do

You want to compare two agents on cost. What do you measure?

The advertised price per million tokens on the model you chose How long each one took from first prompt to finished change Total input tokens for the same task, from a fresh session How many files each one opened, edited and left changed

What is the cheapest habit change available to most people?

Move everything to the cheapest model your provider offers Have the agent skip the test run and check the change yourself Ask shorter questions so the model has less to read each turn Start a new thread whenever the subject genuinely changes

Why does prompt caching matter so much for agents specifically?

A cached model answers from memory, so its judgement improves It shortens the model's replies, which is where the money goes Agent turns resend near-identical prefixes, and cache reads cost less It stores file contents, so the agent stops re-reading the project