Your coding assistant was a seat license. Forty dollars a developer, renewed once a year and rarely questioned. It sat in the software budget beside the design tool and the ticketing system, behaving the way software always had.
Then the invoice stopped being flat, and now the cost of a developer is anyone's guess.
Gartner projects that AI coding spend will pass the average developer salary by 2028, driven by consumption pricing and rising token use. We take no position on that forecast. What we can speak to is what engineering and finance leaders told us across roughly a hundred conversations in the first half of 2026, and they pointed the same way.
The reasonable version of the objection
The pushback we hear is fair. Developers are expensive, the assistants make them meaningfully faster, and a few thousand dollars a month against a loaded engineering salary is obviously worth it.
At one company, the head of product estimated that somewhere between fifty and sixty percent of the code their team ships now comes through Claude. Another customer told us that more than seventy percent of their developers picked Devin over Cursor and Codex when given the choice. A third went all in on Claude across the whole organization. These teams run on it in daily production.
The value is real. The accounting underneath it broke, and most teams have not caught up to that.
The number that stays out of reach
Most engineering leaders cannot say what one developer costs in AI tooling per month. Pressed for a figure, they admit they are estimating.
The CTO at one company was one of the few who had actually run it, and what they found was a gap between the market rate and their own. "$2,000 per person per month is usual across, but my engineers are at around 500 to $600 per person per month." That is a spread of roughly four times between what they expected to pay and what they actually pay, and the only reason they know is that somebody sat down and worked it out.
The spread widens further at the extremes. A partnerships lead at another company described a colleague spending between a hundred fifty and two hundred euros a day on a single agent. That one developer, at that rate, costs more per month than an entire engineering team running in that five-to-six-hundred-dollar range costs per head.
At real headcount, that variance compounds fast. One customer has two hundred thirty-two developers enabled on GitHub Copilot. A software architect at another company put their development-side AI cost in the mid six figures annually and was clear they had not reached seven figures yet. At those headcounts, a four times spread in per-developer cost is not a rounding difference on a software line. It is the difference between a budget that holds and one that does not.
Why the number is hard to get
Three forces work against an accurate number, and staring harder at the invoice fixes none of them.
First, pricing moved underneath the tools. An engineering manager at one company described Cursor shifting to charging per pull request, which mattered because Cursor was doing most of their code review. Seat costs multiply cleanly, while per-PR costs swing with how the team happens to work that week.
Second, the per-user layer is missing from the data. A platform engineer on one company's internal GenAI team described the specific thing they could not see. "The challenge we have had is really getting granular per user to see where the inefficiencies lie. Like the user that defaulted to opus when most of what they [do] could be done with Haiku." That is a routing problem, and it surfaces only at the per-user level, well beneath the account-level bill where the money lands.
Third, the accuracy tradeoff gets made without numbers. A product and engineering exec at another company walked us through the decision they face constantly, where a workflow gets slightly more accurate, and they have no idea whether token usage doubled to get there. "there's always the cost accuracy trade off... and you're often balancing it blind."
The CTO at one company asked us a question that showed how early they still were. "Is there a way that I could use the free tier to at least measure how much my team is actually using so that I know how much this is going to cost?" They needed a baseline for current spend before they could plan any reduction.
The productivity math is shakier
If you can’t state the cost per developer, the fallback is to justify the spend on output. That measurement is in rougher shape than the cost measurement.
A consultant we spoke with warned us about the metric most teams reach for first. "Capturing the number of actual commits in Git repos... I feel like is potentially driving the wrong results because it's just incentivizing creation." Measure commits and teams simply produce more commits, useful or not.
The same CTO who had actually run the numbers had already tried the more sophisticated version. After working with developer experience platforms that infer productivity from pull request data, they concluded the inference was too loose to rely on. "I don't really trust the PR data. Currently I don't really trust the tokens."
Both halves of the ROI calculation are unreliable at once, which is where most engineering organizations sit today. That gap is what gives the Gartner headline its force. The worry is a number that keeps climbing while it stays impossible to defend.
Where we fit
Per-developer attribution comes first, because every other question depends on it.
Revenium's AI Assistants tracks assistant usage across Claude Code, Copilot, Cursor, Gemini CLI, and Codex in one place. The doc, Connect Your Coding Assistant & Code Repository, shows you the setup path to see your own coding assistant numbers before deciding anything. Configure AI Assistant Pricing Mode handles the seat versus consumption question directly, letting you choose whether a team's usage counts as real spend or stays an API-equivalent estimate. You can see the list of every data point collected in the AI Assistant Data Reference doc, to check whether the per-user visibility you need is actually there. Finally, our GitHub Integration brings PR attribution alongside the usage data, with the caveat that PR counts carry the same incentive problem one consultant flagged.
A useful place to begin is the cost of your most expensive developer. That figure usually settles the debate over whether this needs governing.
