The formula
We work out what one task costs - the planning, all the in-between steps, the tool uses and the final answer - then add a bit for the steps that have to retry. Multiply by how many tasks you run a month and add any human review. We also show the 'cost per successful task', which spreads the cost of the failed runs across the ones that worked.
Questions
Why does an agent cost so much more than a chatbot?
A chatbot answers in one go. An agent does a whole job in many steps - plan, use a tool, check, try again - and pays for each one, carrying its notes along every time. One agent task can cost as much as dozens of chatbot replies.
What's 'cost per successful task'?
Some tasks fail, but you still pay for the work they did. This number spreads that wasted cost across the tasks that actually worked, so it's the true cost of one finished result.
How do I make an agent cheaper?
Use fewer steps, send the routine steps to a cheap model, reuse the repeated instructions with caching, and make tools reliable so fewer steps fail.