Who it's for
Engineering leaders who have to budget for AI agents.
You'll be able to
Pick a few numbers to track from day one, and read outside cost figures with care.
Token prices keep falling. Bills keep rising. Count cost per finished task, not per token.
Why the bill grows
The price to reach a fixed level of model capability has fallen fast, year on year. Spend per engineer still rises, because agents use far more tokens than chat. One vendor's documentation says a team of agents in plan mode uses about 7 times the tokens of a standard session. Long contexts and missed caches add more.
So a cheaper token does not mean a cheaper change. The number to track is what one delivered task or pull request cost.
A zero says we counted and found none. An empty value says we do not know.
What other teams spend
Published figures for 2026 disagree, and all of them come from vendors.
| Source | Figure | Who it covers |
|---|---|---|
| Anthropic | 150 to 250 US dollars per developer per month | Average over enterprise users |
| LinearB | 481 US dollars per developer per month | The 90th percentile of spenders |
| Larridin | About 920 US dollars per developer per month | Only engineers who merged code and had billed AI spend |
The gap is mostly about who is counted, not about price. If you put one of these in a budget, say which population it describes. There is no independent survey of per-engineer spend yet, and no standard definition of "cost per merged pull request" or "tokens per task". You will have to define your own.
The numbers to watch
Log one line for every agent run, every gate step and every pull request. From that log you can get:
- Tokens and estimated cost per run, split by model.
- Cost per delivered task, from runs joined to merged pull requests.
- Gate runs per merged change, and gate time. In our demo project, one weekly review counted 4.83 gate runs and 19.2 minutes of gate time per merged change.
- How often a person had to step in during a run.
- Pull requests opened, first reviewed and merged, and which ones an agent wrote.
Keep the log to numbers, times and labels. Never put prompts, answers, secrets or personal data in it.
An estimate is not a bill
Cost worked out from token counts and a dated price list is an estimate. It does not know your discounts, a flat monthly plan, tax, or a price that changed later. Say "estimate" next to every cost figure.
When a value can't be known, record it as unknown with a reason, never as zero. A zero hides the gap and makes averages look better than they are.
Try it
Pick five pull requests your team merged with AI help last month. For each, write down the tokens or cost, the number of gate runs, and the review time. Mark each value you can't find as unknown. Count the unknowns. That count tells you what to log first.