What it costs to run a text-in, text-out agent on your own API key: per run and per month, for three common shapes, priced from published API rates. The platform (n8n, Make, or your own code) bills separately; the model calls are the number below.
one call per item: read a feed entry, write a summary
$0.0067–$0.010 80% of runs
Varies mainly with how long each feed item runs.
Call figures are medians over the same simulated runs as the headline, so they don’t sum exactly to it: no single run lands every range’s high end at once.
a planner call, then ~5 tool steps that each re-send the growing context
$0.063–$0.095 80% of runs
Varies mainly with how many tool steps a run takes and how much each step reads.
Call figures are medians over the same simulated runs as the headline, so they don’t sum exactly to it: no single run lands every range’s high end at once.
a planner, ~20 tool steps, then a frontier writer over ~11k tokens of gathered context
$0.447–$0.660 80% of runs
Varies mainly with how many research steps a run takes, how much gathered context the writer reads and how long the report runs.
Call figures are medians over the same simulated runs as the headline, so they don’t sum exactly to it: no single run lands every range’s high end at once.
How these are priced: from OpenAI’s published per-token rates as of September 19, 2026, with gpt-5.6-terra on the routine calls and gpt-6-astra where an agent plans or writes. OpenAI is the default here because it is where most agent builds start; in the simulator, any call can be re-priced on another provider’s model. Published rates move; the API pricing log keeps a dated record of the changes.
The ranges above come from 2,000 simulated runs in which the step count and the token sizes vary. Three things move a real bill that those ranges only partly cover.
Every tool step re-sends the conversation so far, so input grows with each step. The research agent’s 20 steps, at 1.5k tokens added per step, re-send about 285k tokens over a run. That growth is in the ranges above.
A call that fails still bills its input, and the retry bills it again. Most platforms retry two or three times by default, so one flaky step on the triage agent adds about $0.0077 per attempt, and a run that exhausts its retries pays for three or four attempts and returns nothing. Retries are not in the ranges above.
An agent that doesn’t recognize it’s done keeps stepping, and with context growth the cost scales with the square of the step count, not the count. The research agent’s 20 steps re-send about 285k tokens; 200 steps would re-send about 30M, a hundred times more. Cap the step count in the workflow, and set a spend limit at the provider. Loops are not in the ranges above.
Was this helpful?