How much does an AI agent cost?
Per token, an AI coding agent costs what its model costs: Claude Opus 5.5 is $4 per million input tokens and $20 per million output. Per task, it depends on how you run it. Per month, it depends on how many tasks you hand it.
The numbers below come from our simulation of one agent working one 5-point ticket, “Add cursor pagination to /orders”, run 20,000 times for each setup and priced at list prices.
What one agent run costs
“Carefully” means targeted context, at most 30 cycles and a $15 cap with a check-in at $10. “Wide open” is how most people start: the whole repo in context, no cycle limit, no cap. The game stops a run once it passes its $50 budget, so wide-open runs top out a little above it.
| Model, how it's run | Cheap run | Typical | Costly run | Ships |
|---|---|---|---|---|
| Claude Opus 5.5, carefully | $1.63 | $4.56 | $15.98 | 87% |
| Claude Opus 5.5, wide open | $7.43 | $52.53 | $61.68 | 33% |
| Claude Sonnet 5.5, carefully | $0.80 | $3.80 | $16.28 | 84% |
| Claude Sonnet 5.5, wide open | $6.45 | $52.20 | $56.77 | 32% |
| Claude Haiku 5.5, carefully | $0.11 | $8.96 | $16.19 | 62% |
| Claude Haiku 5.5, wide open | $14.72 | $50.42 | $51.49 | 27% |
Cheap run is the 10th percentile, typical the median, costly the 90th. Ships is the share of runs that finish the ticket before a cap or the budget stops them.
Run carefully, a typical Claude Opus 5.5 run costs $4.56 and 87% of runs ship. Wide open, the typical run costs $52.53 and 67% of runs burn the whole budget without shipping.
Why the same task can cost 30x more
Run the same ticket twice with the same settings and you get two bills. What the agent reads first, a test that fails for an unrelated reason, a loop it can't get out of: each one changes how many cycles it takes, and every cycle is paid for.
Bai, Brynjolfsson, Pentland et al. (arXiv:2604.22750, 2026): runs of the same task can differ by up to 30x in total tokens. arXiv:2604.22750 (v3, 2 Oct 2026)
Our simulation shows the same thing. Across careful Claude Opus 5.5 runs, the most expensive 2% cost 14x the cheapest 2%. Wide open, the gap is 25x, and that is with the $50 budget cutting the top off.
What drives the bill
- Model tier. Claude Opus 5.5 output costs $20 per million tokens, Claude Haiku 5.5 $0.50. The small model needs more cycles to land the same fix: careful Claude Haiku 5.5 runs averaged 10.9 cycles, Claude Opus 5.5 2.2.
- Context. Whatever is in context is sent again on every cycle. Caching makes the repeats cheaper ($0.20 per million on Claude Opus 5.5), and a whole repo in context still adds up fast.
- Task length and retries. A long task means more cycles. A loop that never lands the fix keeps paying until an iteration limit or a cap stops it.
- Sub-agents. Running several in parallel gives more chances to land the fix and multiplies the tokens. Sometimes they overwrite each other.
The method page explains each of these dials in the game.
Cost per developer per month
Multiply by volume. At 2 medium tasks a day for 21 working days, a careful Claude Opus 5.5 agent comes to about $283 per developer a month at API rates. At 6 a day it is $848. Run wide open at 6 a day, it is $5,536.
Two published figures for comparison:
Anthropic's Claude Code docs: across enterprise deployments, the average cost is around $13 per developer per active day and $150–250 per developer per month, with 90% of users below $30 per active day. Claude Code docs, Manage costs
Gartner analyst Nitish Tyagi told The Register that AI coding bills are leaping from $20–100 to $2,000–5,000 per developer per month, with extreme cases reaching $20,000. The Register, 24 Jun 2026
Agent cost vs a subscription
Agent tools sell seats, and a seat doesn't always cap the bill. Some plans add every token on top of the seat price.
| Tool | Seats | How usage is billed |
|---|---|---|
| Claude Code | $20 to $200 a seat | Flat plans with usage limits, an API key billed per token or a seat price plus every token at API rates |
| Cursor | $30 to $200 a seat | A seat price plus every token at API rates |
| GitHub Copilot | $10 to $100 a seat | Seats that include usage credit, then API rates |
A flat plan is cheaper once the agent works most days. Light use is cheaper per token. The break-even point for each plan is on API vs subscription.
Estimate your own agent
Play a ticket in the game and watch the meter, or put your team size into the calculator for Claude Code, Cursor or GitHub Copilot.
Reality check
These are simulated runs, calibrated to look like real agent work. Your codebase, prompts and tool will move the numbers. Treat them as orders of magnitude. Cache-write surcharges, batch discounts and regional pricing are left out. Method and sources.
This page is independent and not affiliated with any model or tool vendor.
See where the money goes, one run at a time.
Play the $50 agent game