Three 2026 datasets on what coding agents actually cost per developer, where the money goes, and the point where spending more stops producing more.

For years a developer tool cost about $20 a seat. You bought the seats, forgot about them, and nobody in finance ever asked.

That era ended this year. Coding agents are billed by the token, and the token bill behaves nothing like a seat licence.

Three separate datasets from 2026 now show what that looks like in practice. They agree on the direction. The interesting part is what they say about where the money goes.

The bill, in three datasets

Jellyfish looked at Claude Code spend across its customer base for its H1 2026 State of AI in Software Engineering report. Between August 2025 and May 2026 the median user went from about $5 a month to $81, a 16x climb in ten months.

Scale Venture Partners asked 38 startups for their monthly code generation spend, seats and tokens together. The answer, published in May: a median of $577 per engineer per month and an average of $1,040. The author puts that at roughly 50 times what a developer tool used to charge per seat.

Gartner went furthest. In June it predicted that AI coding costs will pass the average developer's salary by 2028. Its peer community data, as reported, has roughly a quarter of technology leaders already paying $200 to $500 per developer per month for tokens, and about 6% paying more than $2,000.

That Gartner headline needs one footnote, and Gartner's own analyst supplied it: the prediction uses a global average salary of about $2,000 a month. It is not a claim about a San Francisco pay packet.

The average is the wrong number

Notice that in the Scale data the average is nearly double the median. That gap is the real story.

Jellyfish's May 2026 distribution shows it clearly. More than half of users still spent under $100 that month. The mean was $248. The 99th percentile was $2,452.

Token use follows the same shape. The median Claude Code user's weekly token count stayed flat through 2026, while the 90th percentile tripled, from about 50 million tokens a week to 170 million.

So most of your team costs about what a nice lunch costs. A handful cost what a contractor costs. A per-seat budget built on the average will be wrong for almost everyone on it.

Where the money actually goes

Here is the number that changes the picture.

In Jellyfish's data, generated code is 0.4% of all tokens. Cache reads, the agent re-reading context it has already seen, are 96%.

Cached tokens are billed at a fraction of the normal rate, so the bill is less lopsided than the token count. But it is still lopsided: cache reads and writes make up about 83% of the bill, against 16% for generated output.

You are not paying the agent to write. You are paying it to read. Every turn of a long session carries the whole conversation, every file it opened, and every tool result back through the model again.

That reframes the cost problem. The expensive thing is not how much code you ask for. It is how much context you drag along while asking.

More tokens, same output

If the heavy spenders were shipping proportionally more, none of this would matter. Jellyfish checked, using Cursor data from more than 300 companies.

  • The first 0.7 million tokens per user return almost 11 pull requests per million tokens
  • Returns diminish after about 2 million tokens per user per month
  • Between the third and fourth spending bands, the marginal return is negative

Output does tick up again in the very highest band, and the report works out when that is worth it: heavy spend pays off once a merged PR is worth about $69 or more to you. Value a PR at $15 and the top band's net value drops below zero.

Gartner's analyst said the same thing in fewer words: maximising token use "is not directly related to higher productivity gains."

What to do about it

Budget by distribution, not by seat. Look at your own p50, p90 and p99. Set an alert on the tail rather than a cap on everyone.

Keep sessions short. Since the bill is mostly re-read context, a fresh session for a new task is the cheapest optimisation there is. A four-hour session pays for its first hour over and over.

Invest in context files, not context volume. Jellyfish found that each doubling of a team's investment in context files bought roughly 29% more merged PRs per developer per week. A short, accurate project guide beats letting the agent rediscover the codebase every morning.

Route by difficulty. Renames, boilerplate and test scaffolding do not need the frontier model. This is Gartner's recommendation too: small models for frequent simple work, the expensive one for the hard problems.

Measure cost per merged PR. Tokens per developer tells you who is busy. Cost per merged change tells you who is effective.

One caveat

Jellyfish sells engineering analytics, Gartner sells advice, and the Scale sample is 38 startups from one investor's network. None of them is a neutral census. The Jellyfish spend figures are also specific to its customers and mostly to Claude Code. But three different methods landing on the same skewed curve is hard to dismiss.

The takeaway

The bill is not growing because agents write more code. Generated code is a rounding error on the invoice.

It is growing because agents read more: longer sessions, bigger contexts, more of them running at once. That is good news, because reading is the part you control.

The cheapest token is the one the agent never had to re-read.