Part 2 of 12 · 2 min read
Cost per resolved task, not cost per call, is the number that matters
A task the agent attempts and fails still cost money — the unit that matters is the one that actually got done.
Think of a delivery company that only counts the cost of trucks that arrive. It ignores the ones that broke down halfway. It would look far more efficient than it actually is.
In plain words: cost per model call looks cheap, but a task the agent escalates or fails still cost something to attempt, so the honest number is cost per task divided by the share of tasks the agent actually resolved on its own.
A worked example. For the support task (all inputs assumed): model tokens cost $0.062, tools $0.010, infrastructure $0.005, and credits for failed tasks $0.150 — a total of $0.227 per task attempted. Dividing by the 55 percent the agent resolves alone gives $0.41 per agent-resolved task. For a heavier research task, the same arithmetic gives $4.725 per task attempted and $5.91 per resolved task — and here 77 percent of that cost is model tokens, not credits, which changes which lever actually matters (covered next).
The key lines. From code/mini-pricing/economics.py:
# from code/mini-pricing/economics.py
def cost_per_task(a, k=1.0):
"""What the vendor spends per task: model, tools, infrastructure and credits for failed tasks.
Escalated tasks are handled by the customer's own people, so their cost is the customer's, not the vendor's."""
pa, pe, pf = outcomes(a, k)
return model_cost(a, k) + a.tool_cost + a.infra_cost + pf * a.fail_credit
def cost_per_resolved(a, k=1.0):
"""Vendor cost per task the agent resolved by itself: failures and escalations still cost money."""
return cost_per_task(a, k) / outcomes(a, k)[0]
Read it as: cost_per_resolved divides the full attempted-task cost by only the resolved fraction. A lower resolution rate doesn't just mean fewer successes. It means every dollar spent on attempts gets divided among fewer of them. That is why this ratio moves faster than the raw per-task cost does.
Unit economics, cost per resolved task
Unit economics is the revenue and cost for one unit of the product — here, one task or one customer. Cost per resolved task is the total cost of every task attempted, divided by the number the agent actually resolved.
Example: the $0.227-to-$0.41 jump above is exactly this division in action — the 55 percent resolution rate is the denominator that turns the smaller number into the larger one.