AI Agent Cost Estimator
See what one agent run actually costs, how many single API calls that adds up to, and which step the money disappears into.
Model calls to finish one task.
Completed tasks, not API calls.
Every step re-sends everything before it. The default in most frameworks.
Refine the inputsHide the detail
Repeats on every step.
Tool results, user turn.
Costs 4x to 5x input.
Steps that fail and re-send the same context.
Tool selection, routing, extraction.
Where the run gets expensive
Input tokens per step. The dark band is your system prompt, the bright band is conversation being re-sent, and the thin band at the end is the only genuinely new information in that step.
What each control is worth
Each figure is this model re-run with that one thing changed, so it is a real delta rather than a rule of thumb. They are not additive.
- Cap iterations at 6saves $6,251/mo
Half the steps. The cut is more than half the cost, because the steps you remove are the largest ones.
- Route half the steps to a small modelsaves $4,759/mo
Tool selection and routing rarely need the frontier tier.
- Cache the system promptsaves $2,663/mo
The prefix repeats on every step, so it is the most cacheable span in the run.
- Summarise instead of carrying everythingsaves $2,263/mo
A running summary plus the last few turns, rather than the whole transcript.
- Remove the retry surchargesaves $1,307/mo
A circuit breaker on repeated identical failures. Retries bill in full and log once.
Why agents cost what they do
Three things drive the multiple. The estimator models all three; here is what each one is doing to your bill.
Context compounds
Every step re-sends the conversation so far, so input tokens grow with roughly the square of the step count. This is why the multiple gets ugly fast.
Retries bill in full
A step that fails and retries re-sends the same large context and is charged again, while most logging records the outcome once.
The last steps are the dear ones
Cutting a run from twelve steps to six removes much more than half the cost, because the steps you remove carried the most context.
The full write-up is in AI agent costs. To work out which cost lever applies to your whole system rather than one agent, use the AI cost diagnostic.
AI agent cost FAQ
How much does it cost to run an AI agent?+
Far more than the same task as a single completion, because an agent does not send one request, it sends one per step and each step re-sends everything before it. A twelve-step run on a frontier model with a 4,000-token system prompt typically lands around 20x to 30x a single call. Push it to thirty steps and the multiple passes 100x. The step count and what you carry between steps decide it, not the model price.
Why do AI agents cost 19x to 50x a single API call?+
Because context accumulates. Step twelve carries the system prompt, the tool definitions, and every prior thought and tool result, so total input tokens grow with roughly the square of the step count rather than in a straight line. Retries multiply on top: a step that fails and is retried re-sends the same large context and is billed again. This estimator computes the multiple for your own numbers instead of quoting the range.
What is the biggest lever on agent cost?+
Almost always the step count, because the steps you remove are the most expensive ones. Halving a twelve-step run to six removes far more than half the cost, since steps seven through twelve each carried more context than any step before them. A hard iteration cap is usually worth more than any model swap, and it is a smaller change.
Does caching the system prompt help an agent?+
Yes, and agents are the best case for it. The system prompt and tool definitions repeat on every single step of every run, so they are the most re-read span you own. Cached tokens bill at roughly a tenth of standard input price. The saving grows with step count, which is the opposite of most optimisations, and it applies whether or not you also fix the loop.
How do I reduce the context an agent carries?+
Stop sending the full transcript. Keep the last few turns verbatim, replace everything older with a running summary, and drop tool results once they have been consumed. Switch this tool between "full transcript" and "summary plus last N" to price the change on your own run before you build it.
Are multi-agent systems more expensive than one agent?+
Usually yes, and by more than the agent count suggests. Each agent carries its own system prompt and tool definitions, and handoffs mean context gets re-sent across boundaries rather than staying in one accumulating thread. Model a multi-agent system by summing the runs, then add the handoff payloads as fresh tokens on the receiving side.
Agents that survive their own invoice
The estimator shows the multiple. We build the iteration caps, circuit breakers and context discipline that bring it down. Book a free cost review.