TokenPad

Cost scenario · Automation

An agent that uses tools

An agent that researches a question using tools: it searches, reads results, decides whether it has enough, and either continues or answers. A typical task runs five iterations; the cap is twelve.

The number that matters is not the average. It is the worst case — cost per iteration multiplied by the cap — because that is what appears during an incident, and an agent without a cap has no worst case at all.

The shape of one task

38,000input tokens
3,000output tokens
35%of input is cacheable
5model calls per task

A 1,500 token system prompt with tool schemas, constant across iterations, plus a trace that grows by roughly 1,500 tokens each step as tool results accumulate. Summed across five iterations that is about 38,000 input and 3,000 output tokens per task.

What it costs per month

Computed from the verified rates in the price table, last reviewed August 5, 2026. Figures assume caching is enabled — the uncached column shows what skipping that flag costs.

Monthly cost by volume and model for An agent that uses tools
VolumeMinistral 3 8BCheapest that fitsMagistral MediumMid tierClaude Fable 5Frontier
Internal use500 tasks / month$3.08$3.08 uncached$45.50$45.50 uncached$205.15$265.00 uncached
Product feature20,000 tasks / month$123.00$123.00 uncached$1,820.00$1,820.00 uncached$8,206.00$10,600.00 uncached
At scale200,000 tasks / month$1,230.00$1,230.00 uncached$18,200.00$18,200.00 uncached$82,060.00$106,000.00 uncached

On this shape the spread between the cheapest model that fits and the frontier one is 86.2×, which is the largest single decision on this page. Enabling prompt caching removes a further 23% on Claude Fable 5, because 35% of the input on this workload is a stable prefix.

Where a provider publishes no cached input rate, the cached and uncached figures above are identical. That is a gap in what the provider discloses, not a model that cannot cache.

What actually drives the bill

Context accumulates every iteration

Iteration five sends everything from iterations one through four. The tenth iteration of a long task can cost five times the first, which is why average cost per iteration is a misleading number to budget from.

The iteration cap is the real budget

Worst case is per-iteration cost times the cap. An agent capped at twelve costs roughly two and a half times the typical five-iteration task when it goes wrong — and it goes wrong on exactly the inputs you did not anticipate.

Tool results are input tokens

A search returning five results with snippets is a few thousand tokens entering the context and staying there for every subsequent iteration. Verbose tools are expensive tools.

Stuck loops

The characteristic agent failure is a two-step cycle repeated to the cap. Without a cap it repeats until something else stops it, and the something else is usually the invoice.

How to cut it, in order of return

  1. Trim tool results before they enter context30–50%

    Return the three fields the agent needs, not the whole API response. Every unnecessary token is re-sent on every subsequent iteration, so the saving compounds across the loop.

  2. Cache the system prompt and tool schemas25–30%

    Tool definitions are large and completely static. They are the most cacheable content in any agent and are re-sent on every single iteration.

  3. Lower the iteration capdirectly on worst case

    Measure how many iterations real successful tasks take. If the answer is four and your cap is twenty, you are budgeting for sixteen iterations that never help.

  4. Compact the trace mid-run20–40% on long tasks

    Replace early tool results with a summary of what was learned once the trace gets long. It flattens the growth curve exactly where it hurts most.

What people get wrong budgeting this

  • Budgeting from the average task ignores that the expensive tasks are the ones that fail, and failures are not rare.
  • Reasoning models multiply this workload. Reasoning tokens are billed at output rate on every iteration and never appear in what you see.
  • Parallel subagents multiply cost by their count, and each carries its own copy of the context it was handed.

Price your own version

The shape above is representative, not measured from your system. Put your real prompt and volume through these and the number becomes yours.

Before committing to volume, the cost control checklist is ordered by saving per hour of effort, and every rate used here links to the provider page it came from on the Mistral AI and other provider pages.