LEARN / AGENT COSTS

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

PUBLISHED 2026-07-10 · TOKENBURN INDEX

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, with numbers, and the levers that actually move them.

What changes when a chat becomes an agent

A chat is one transaction: context in, answer out. An agent is a loop: read the context, reason about what to do, call a tool (search, run code, edit a file), read the tool's output, reason again — repeat until done. Three billing consequences follow. Every iteration is a fresh model call. Every model call re-sends the entire accumulated context — instructions, history, and every tool result so far — as input tokens. And every iteration generates output tokens you mostly never read: internal reasoning, tool-call formatting, progress narration. The loop is the product. The loop is also the bill.

A worked example, iteration by iteration

Take an unremarkable 20-iteration coding-agent run on a $2-input / $10-output per-MTok model (Claude Sonnet 5, today's price). Context starts at ~3,000 tokens (instructions + task) and grows ~1,500 tokens per step as tool results accumulate; each step emits ~800 output tokens. Watch the cumulative cost compound:

The same question answered in a single chat turn: ~$0.02. That's a ~40x gap with nothing going wrong. Notice the shape, too: the second half of the run costs three times the first half, because context re-reads grow with the square of run length. Long runs aren't linearly expensive — they accelerate.

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, so input cost grows quadratically with run length. An agent that “just needs a few more steps” is quietly doubling its bill.
• — Invisible output. Reasoning tokens, tool-call JSON, and progress narration all bill at output rates — 4–8x input. You see the final answer; the meter saw everything.
• — Failure is billed at full fare. Attempt #23 of a retry loop looks exactly like attempt #4 to the meter. Most agent-cost horror stories are failure-mode stories.

The failure gallery

The expensive patterns are predictable enough that we rank them in the Burn Index: the coding agent stuck in a retry loop on a failing test (every attempt re-reads the repo); the research agent with unbounded browsing (it read the whole internet, it cited two tweets); the multi-agent “crew” whose reflection step has five models debating each other at output rates (conclusion: “it depends”); and the agent that re-reads the entire repository every task because nobody scoped its context. None of these are exotic. All of them are budget line-items somewhere right now.

The six levers that actually work

• — Cap iterations. A hard stop turns a runaway loop into a bounded cost. Boring, effective, non-negotiable. Set it before you need it.
• — Cache the stable prefix. Instructions, tool definitions and base context re-sent every step are exactly what prompt caching discounts by ~90%. For agents, this is usually the biggest single saving.
• — Scope the context. Send the relevant excerpt, not the whole repository. Prune or summarize old tool results instead of dragging them through every remaining iteration.
• — Ask for diffs, not rewrites. Output is the expensive direction. A full-file rewrite pays output rates for every unchanged line.
• — Route sub-tasks to cheaper models. Classification, extraction and summarization steps inside the loop don't need the flagship. A 10–30x per-token spread is sitting there.
• — Watch the meter. Log tokens per run, alert on outliers. An agent bill you look at weekly cannot surprise you monthly.

So are agents worth it?

Frequently, yes — and that's the honest twist. A $0.85 run that replaces twenty minutes of human work is one of the best trades in software. A $40 overnight run that produced a renamed variable is not. The difference is rarely the agent and usually the supervision: bounded loops, scoped tasks, and someone who has actually seen the number. The problem is not the price. It's that the price is quiet.

How to budget an agent workflow

Take your agent's system prompt and typical task, and run them through the Burnmeter in agent mode — it applies the loop multiplier to estimate real per-run cost. Multiply by your expected monthly runs. If the number makes you flinch, revisit the levers above before revisiting the idea. Your agent did not work harder. It just spent more.

Frequently asked questions