How much does an AI agent cost?
An AI agent doesn't make one model call per task — it makes several, each one resending everything it has done so far. Enter your tasks per month, steps per task, and tool usage to get a Low / Likely / High monthly AI agent cost estimate. Everything runs locally in your browser — no data is sent to any server.
- No login
- No credit card
- No API calls
- Estimate only
Configure your AI agent
Estimated monthly cost
T9 estimates LLM inference/token cost only. Not included: search APIs, browsing APIs, other external tools, vector databases, hosting, infrastructure, or any other third-party service fee — those are billed separately by whatever tools and services your agent calls.
This is an estimate, not a billing guarantee. Actual costs depend on your provider's exact tokeniser, your agent's real step count and tool usage, and your billing tier.
How this estimate works
- Instruction / system prompt
- The original task
Same instruction and task, resent — plus every prior step's output and tool-call results, now part of the agent's context.
Every step is billed at its own token count and its own applicable rate, then summed into one task's cost.
An "agent" here means a model that calls itself repeatedly to complete one task — reading its own prior output (and any tool results) at each step before deciding what to do next. This calculator prices the full sequence of calls one task takes, not just one call.
Why more steps increase cost
Each step resends the agent's instructions and the original task — that part stays flat. But it also resends everything from every earlier step, so a task's last step is always its most expensive: step 5 carries the accumulated output of steps 1–4 on top of its own work. Cost per task grows faster than the step count itself, because later steps are larger requests than earlier ones.
Why accumulated context increases cost
Every token the agent has already produced — its own reasoning and answers, plus anything a tool returned — has to be resent on the next call, because the model has no memory between API requests. A 10-step task doesn't cost 10× a 1-step task; it costs more, because steps 2 through 10 are each carrying more history than the one before it.
How tool calls affect token usage
Every tool call an agent makes — a search, a function call, a lookup — returns a result that gets added to context for every following step, the same way the model's own output does. More tool calls per step, or tools that return larger results, both raise every subsequent step's input size and, with it, the cost of the rest of the task.
How model choice affects cost
Because an agent resends its whole history on every step, input-token price matters more here than in a single-shot workload — and a long-running, tool-heavy agent can cross a model's long-context pricing tier partway through a task even if the first step doesn't. This calculator applies each model's own current input/output rate, and its long-context tier if it has one, to every single step.
What's not included: the tools themselves
T9 estimates LLM inference/token cost only. Not included: search APIs, browsing APIs, other external tools, vector databases, hosting, infrastructure, or any other third-party service fee.
Those are billed separately by whatever search API, browsing tool, code sandbox, or other service your agent calls — this estimate covers only the language model's own input and output tokens.
Low / Likely / High explained
Minimal instructions, a short task, short step output, light or no tool overhead.
Standard instructions and typical task/step-output length. Plan against this number.
Detailed instructions, a long task, long step output, heavy tool overhead. Budget against this.
Cut your cost
Example monthly AI agent costs
These worked examples use the same default assumptions as the calculator above — standard instructions, medium task and step-output length, one standard tool call per step, 5 steps per task — with GPT-5.4 Mini. Only the number of tasks per month changes between rows. Use the calculator above to model your own usage or a different model.
| Usage level | Tasks / month | Low | Likely | High |
|---|---|---|---|---|
| Light usage | 50 | $0.3863 | $0.6188 | $0.9375 |
| Typical usage | 200 | $1.55 | $2.48 | $3.75 |
| Heavy usage | 1,000 | $7.73 | $12.38 | $18.75 |
And here's the same default task, with everything else held constant, at increasing step counts — showing why cost per task doesn't scale in a straight line with steps:
| Steps per task | Low | Likely | High |
|---|---|---|---|
| 1 | $0.001065 | $0.001650 | $0.002475 |
| 3 | $0.003915 | $0.006188 | $0.009337 |
| 5 | $0.007725 | $0.0124 | $0.0187 |
| 10 | $0.0215 | $0.0351 | $0.0534 |
Caveats
- These estimates use the same reference token ranges as the site's chatbot calculator for instruction and task/step-output length, accurate to ±10–15% for plain English text.
- Tool-call/tool-result overhead sizes ("light"/"standard"/"heavy") are this project's own descriptive assumption, not a provider-published figure — no AI provider publishes a token-overhead spec for tool calling, since it depends entirely on the specific tool. Real overhead varies widely by tool; use your own measured token counts where you have them for the most accurate estimate.
- Each provider uses a different tokeniser. Actual token counts may differ ±5–10% between providers for the same content.
- The tools an agent calls are not priced here. Search APIs, browsing APIs, other external tools, vector databases, hosting, and infrastructure are billed separately by whatever services you use — this tool only estimates the language model's own input/output token cost.
- Prices shown are manually verified against each provider's official pricing page as of 2026-09-08. AI provider pricing changes frequently — verify at your provider's official pricing page before making budget decisions.
- The long-context pricing tier, where a selected model has one, is resolved separately for every step of a task (not just once for the whole task) — a later step can cross the threshold even when the first step doesn't, and this estimate applies the higher rate to exactly the step(s) that do.
- The context-window warning above appears when a task's largest single step's combined tokens exceed 80% of the selected model's context window — a safety margin before truncation or failure risk, not a hard limit at 100%.
This is an estimate, not a billing guarantee. Always confirm against your provider's usage dashboard and official pricing before committing a budget.