Agent cost simulator
A single chat reply is one call. A real agent plans, calls tools, checks its work, and retries, and it drags its context along on every step. This estimates the cost of a finished task, which is the only unit that matters once you ship.
The loop
Pricing & volume
- Effective calls / task
- 0
- Tokens / task
- 0
- Cost / finished task
- $0
- Per 1,000 tasks
- $0
- Monthly (30 days)
- $0
Estimate only. The numbers you enter are the model. Defaults are seeded with GLM-5.2 pricing.
Why iterations are the budget
The single biggest cost driver is the one people leave out: how many times the loop runs. Carried context multiplies that, because the same system prompt and history are re-sent on every step. A model that is cheaper per token but needs more steps or retries can easily cost more per finished task. This is the practical heart of loop engineering and cost per outcome. To check whether a smaller fine-tuned model would pay for itself, see reinforcement fine-tuning.