What an AI agent actually costs to run
Everyone quotes a build price. Almost nobody quotes the running cost, and the arithmetic is not obvious — the conversation you resend every turn is where the money goes. Here is the full calculation, including the caching setting that silently does nothing.
A proposal for an AI agent usually has a build price and a go-live date. It rarely has the number the finance director actually asks for three months later: what does this cost per month, forever?
That number is calculable before you build anything. Here is how.
The part people get wrong
The intuitive model is "messages × price per message". That undercounts, often by a factor of three or four, because a language model has no memory between calls. Every turn resends the entire conversation so far.
Turn one sends the system prompt plus one question. Turn six sends the system prompt plus five previous questions plus five previous answers plus the new question. The input grows on every turn while the output stays roughly flat.
For a six-turn conversation with a 600-token system prompt, roughly 60 tokens per question and 250 per answer, the input side totals about 8,600 tokens — not the ~4,000 a per-message estimate suggests.
Running the numbers
Using Anthropic list prices as of July 2026, for that same six-turn conversation at 2,000 conversations a month:
| Model | Per conversation | Per month |
|---|---|---|
| Haiku 4.5 | ~$0.016 | ~$32 |
| Sonnet 5 | ~$0.048 | ~$97 |
| Opus 5 | ~$0.068 | ~$136 |
Two things follow immediately.
The model choice is most of the decision. Not the prompt, not the architecture — the model. A support agent answering scripted questions from a short system prompt does not need the expensive one, and the gap is 4x.
The absolute numbers are small. A support agent handling 2,000 conversations a month costs less than most teams' coffee budget. The risk in these projects has never really been the token bill; it is building the wrong thing, or building something nobody can maintain. Be suspicious of any proposal where inference cost is presented as the main commercial consideration.
The setting that silently does nothing
Prompt caching lets a model re-read a fixed prefix — your system prompt, your tool definitions — at roughly a tenth of the normal input rate. It is the first optimisation everyone reaches for.
It has a minimum. If your cacheable prefix is shorter than the model's floor, the cache marker is simply ignored. No error. No warning. No saving.
Those floors differ per model, and they are not intuitive:
- Opus 5 — 512 tokens
- Sonnet 5 — 1,024 tokens
- Haiku 4.5 — 4,096 tokens
So a 600-token system prompt caches on Opus, and does nothing at all on Haiku or Sonnet. Teams routinely ship caching, see no change in the bill, and conclude caching does not work. It works; the prefix was too short to qualify.
If caching matters to your economics, the fix is usually to move more stable content into the cached prefix — tool definitions, retrieval instructions, few-shot examples — until it clears the floor. That is a design decision, not a config flag.
What actually moves the number
In rough order of impact:
- Model selection. Route by task. Most agents have one hard path and several easy ones; they do not all need the same model.
- Conversation length. Because history compounds, an agent that resolves in four turns costs far less than one that resolves in ten — and it is a better experience anyway.
- Context discipline. Everything you attach is paid for on every subsequent turn. Retrieval that returns ten documents when two would do is a recurring cost, not a one-off.
- Caching, once the prefix is genuinely over the floor.
Work it out for yourself
We put the whole calculation behind sliders on our AI engineering page — move it to your own volume and conversation length and watch the models diverge. The assumptions are printed underneath it, because a number you cannot interrogate is not worth much.
None of this replaces measuring your real traffic. It does mean the order of magnitude should never be a surprise, and that anyone quoting you an agent without a per-interaction cost has not finished the engineering.