What Agentic Workloads Actually Cost
AI pricing guide · updated 2026-08-07
An agent turns one user request into many model calls, and each one generates. Because output costs at least twice what input does on most models, that multiplication lands on the expensive side of the price sheet. Agentic cost surprises are almost always a volume story rather than a rate story.
One request, many calls
A single task becomes a plan, a tool call, an interpretation of the result, a revised plan, another tool call, and eventually an answer. Ten model calls for one user action is unremarkable.
Each call also carries the accumulated history: the original instruction, the steps taken and the results so far. Input grows through the loop while output is generated at every step, so cost scales worse than linearly with task complexity.
Output is the expensive direction
The median model in our catalogue charges twice as much per output token as per input, and 117 of the 271 we track charge at least three times as much. Since agents are output-generating machines, an agent workload sits squarely on the wrong side of that ratio.
This is why the standard blended figure understates agent costs. Weighting output at 30 per cent describes a chat exchange, not a loop that writes at every step. Our rankings publish an output-heavy weighting for exactly this reason.
Reasoning tokens you pay for but never read
Models that reason before answering generate an internal chain of thought, billed at the output rate, often hidden from the response. On a hard problem that can dwarf the visible answer.
Used deliberately on genuinely difficult steps, that is money well spent. Applied to every step in a loop, including formatting and simple tool selection, it multiplies the most expensive token type across the whole workload.
Controlling the bill without crippling the agent
Route by difficulty. Send planning and final synthesis to a strong model; handle parsing, formatting and routine tool selection with a cheap one. Intermediate steps are usually most of the calls, so this captures most of the saving.
Cap the loop. A maximum step count and a maximum output length per step turn an unbounded worst case into a bounded one. Without a cap, a single confused run can cost more than a thousand successful ones.
Reserve reasoning mode for steps that need it rather than enabling it globally.
Summarise history instead of resending every step verbatim, and cache the stable instruction prefix, which every call in the loop shares.
Finally, log tokens per completed task rather than per call. Per-call cost looks reassuring and hides the multiplication; cost per finished task is the number that predicts the invoice.
Frequently asked questions
Why are AI agents expensive to run?
Because one user request becomes many model calls, and every call generates output, the costliest token type on most price sheets. The median model we track charges twice as much for output as for input, and 117 of 271 charge at least three times as much. Accumulated history also grows the input side through the loop.
How do I reduce agent costs?
Route routine intermediate steps to a cheaper model and reserve the strong one for planning and final synthesis, since intermediate steps are most of the calls. Cap the step count and the output length per step, enable reasoning mode only where it is needed, summarise history rather than resending it, and cache the shared instruction prefix.
Do I pay for a reasoning model's hidden thinking?
Yes. Intermediate reasoning is billed at the output rate even when it never appears in the response, which is why reasoning models can cost several times more than their headline rate suggests for a given task. Enable it per step rather than globally.
What is the right metric for agent cost?
Cost per completed task, not per call. Per-call figures look small and conceal the multiplication across a loop, while cost per finished task is what actually scales with usage and predicts the bill.
See it in the data
Related guides
Terms used in this guide
Published by Tokenando. Last updated 2026-08-07. Figures in this guide are computed from our own pricing index and dated where they can move; see the methodology and corrections policy.