Cheap tokens can hide an expensive agent: count cost per successful task

legendpizzasenpai · reddit · 2026-10-15

A Reddit post argues that when comparing agent workflows, cost per successful task is a better denominator than cost per model call — a cheap call can still trigger multiple searches, retries, and results you reject.

The proposed method:

Log billed model calls, tool calls, and compute for every attempt, including failures (use the provider's actual usage records)
Define an acceptable output up front (e.g., a brief with sources passing manual review) — an HTTP 200 is not a completed task
Divide total billed cost by accepted outputs, including fees

Illustration: 100 attempts costing $15 with 75 accepted outputs = $0.20 per output; with a 10% fee on purchased credit, cash cost is $16.50, or $0.22 each. Unused prepaid balances should be tracked separately from run costs. Hold task and acceptance criteria constant when comparing — a cheaper model with more rejected outputs can be the pricier workflow. The author is building DreamTune Labs, a shared-balance service, around this problem.

Original post →

Cheap tokens can hide an expensive agent: count cost per successful task

More from coding & agent