LEARN / AGENT COSTS

Why your AI agent costs 10–40x more than a chat

A chat answers you once. An agent thinks out loud, calls tools, reads the results, thinks again — and every lap of that loop is billed. Agents aren't overpriced; they're doing structurally more metered work than anyone eyeballs. Here's the anatomy.

The loop is the bill

One user request to an agent typically becomes 10–40 model calls: plan, call a tool, read the output, re-plan, repeat. Each iteration re-sends the growing context (instructions, history, tool results) as input tokens and generates reasoning you may never read as output tokens. A task that costs one prompt in a chat window costs a loop as an agent — and the loop compounds.

Worked example

Take a modest 20-iteration coding-agent run on a $3/$15-per-MTok model. Context starts at ~3,000 tokens and grows ~1,500 per step as tool results pile up, so the model re-reads on average ~17,000 tokens per iteration — ~340K input tokens across the run. Add ~800 output tokens per step (~16K total). List price: ~$1.26 per run. The same question answered in a single chat turn: ~$0.03. That's a 40x gap without anything going wrong — and a retry loop on a failing test bills every attempt at full fare.

Illustrative arithmetic at list prices, no caching. Real runs vary wildly — that's rather the point: measure yours.

The three hidden multipliers

• — Context re-reads. The transcript is re-sent every iteration. Cost grows roughly with the square of run length, not linearly.
• — Invisible output. Reasoning tokens, tool-call formatting, verbose progress narration — output rates are 4–8x input rates, and agents produce a lot of it.
• — Failure is billed. Attempt #23 of the retry loop looks exactly like attempt #4 to the meter. Agent burn risk is mostly failure-mode risk.

The levers that actually work

• — Cap iterations. A hard stop turns a runaway loop into a bounded cost. Boring, effective, non-negotiable.
• — Cache the stable prefix. Instructions and repo context re-sent every step are exactly what prompt caching discounts by ~90%.
• — Scope the context. Send the relevant excerpt, not the whole repository. Prune old tool results instead of dragging them along.
• — Ask for diffs, not rewrites. Output is the expensive direction; a full-file rewrite pays output rates for unchanged lines.
• — Route sub-tasks to cheaper models. Classification and summarization steps don't need the flagship.

Are agents worth it?

Often, yes — a $1.26 run that replaces twenty minutes of human work is a bargain. The problem is not the price; it's that the price is quiet. Nobody watches an agent think. Put your agent's typical instructions through the Burnmeter in agent mode (it applies the loop multiplier), multiply by your monthly run count, and decide with the number in front of you. Your agent did not work harder. It just spent more.