Gartner now predicts that by 2028, AI coding costs will surpass the average developer's salary.
By 2028 the tool could cost more than the person using it. That's the line that made me take this seriously.
It is already happening in places. Uber burned through its entire 2026 AI coding budget by April, four months in. Coding agent adoption there went from 32% to 84% of a roughly five-thousand-engineer org in about a month. The tools worked, engineers wanted them, and the bill moved with the usage. Microsoft ended its coding agent licences six months into a pilot. Both were reported publicly, and neither is a story about a badly run engineering team.
A tool your engineers genuinely love can eat a year of budget in a quarter, and almost nothing in the way we buy software is built to notice.
For a few months I've been asking engineering leaders and finance leads a version of the same question: who owns this number? I still haven't had a clean answer.
That's why I'm writing. It's a radically new practice, and it's badly misunderstood.
The misunderstanding: this isn't a licence, it's a meter
Nearly everyone I talk to files coding agents under software licences. A per-seat cost, negotiated once, predictable for twelve months. That instinct is reasonable, and it's the source of almost every mistake that follows.
The vendors moved, and they moved fast. The flat-fee era ended inside about a year.
Cursor went from request-based to credit-based in June 2025. OpenAI moved Codex from per-message to token-based pricing in April 2026. GitHub Copilot completed its move to token-based AI Credits on 1 June 2026. As DX put it after tracking 400-plus organisations over fourteen months: the flat-fee era for AI tooling is over.
Which means the seat price now tells you almost nothing. Consumption tracks how each engineer works on a given day. Someone running an agent for three hours against a large monorepo consumes orders of magnitude more than someone accepting autocomplete suggestions. Same seat, same licence, wildly different cost. DX documented one developer's bill going from $29 to $750 a month, and another from $50 to $3,000.
Multiply that across a few hundred engineers and Gartner's 2028 prediction stops reading like a forecast. Token consumption and consumption-based licensing are the two things it named, and both are already here.
The part I think actually causes the overruns
Here's where I've changed my mind while writing this. I started out thinking the overruns were a measurement problem. I now think they're an ownership problem, and that measurement is downstream of it.
Ask an organisation who owns AI spend and you get a split answer. In DoiT's February 2026 survey of 500 finance leaders, 55% said accountability sits with technology. 53% said it sits with finance. DoiT's own summary is that accountability is "split almost evenly" between the two.
It is tempting to add those together, announce 108%, and conclude that nobody owns it. I nearly did. But the two figures overlapping just tells you some organisations named both functions, and naming both can be a deliberate, well-run arrangement. The finding is narrower and still damning: there is no single function with a clear mandate. Two candidates, each claimed by about half, and no consensus.
The symptom turns up in the same survey. 36% of those leaders named lack of clear financial attribution as a core barrier, which is what you would expect when the mandate is ambiguous.
The reason I find this convincing rather than just tidy is what happens to overruns as organisations get better at measuring. In the same data, the most FinOps-mature organisations had a higher overrun rate than early-stage ones: 89% against 69%. If overruns were fundamentally a measurement problem, maturity would fix them. It doesn't. Mature teams are simply better at seeing the overrun they were already having.
KPMG's Q2 2026 survey lands in the same place from a different angle: 66% of organisations maintain AI cost monitoring dashboards, and only 36% have any direct control over tokens or usage. Two-thirds are watching. A third can act.
Canva is worth a mention here, though not as evidence of the ownership gap. Its problem was the cost of serving AI features to customers, not coding agents. But CEO Melanie Perkins's explanation rhymes with all of this: pricing, consumption model and usage controls "had not caught up with the outsized demand." The controls lagged the adoption. That's the same structural failure, one layer up the stack.
So what would I call it?
Here's where I've landed, for now. Coding agent cost management is the practice of measuring, attributing, and optimizing the spend your AI coding tools generate across an engineering org. Claude Code, Cursor, Copilot, Codex, whatever your developers actually reach for. It treats that usage as a cost someone is accountable for rather than a rounding error nobody is.
A year ago I'd have felt slightly ridiculous writing that. It would have sounded like inventing a discipline in order to sell something. Now it describes a real and growing line in the budget, which is why I think it has earned a name.
Why it needs its own name
Three things make it genuinely different from the cloud cost work many teams already do well.
The billing has no handles. Usage-priced, token-metered, tied to human behaviour. No instance to right-size, no reservation to commit to, no idle resource to switch off. The levers that made cloud cost tractable don't exist here.
The spend is spread across vendors. Most orgs don't standardise on one agent. DX's pattern across those 400-plus organisations is the common one: license Copilot, add Cursor for the power users, roll out Claude Code for senior engineers. That stacks to $200-600 per developer per month across tools, each vendor with its own reporting, its own units, and its own idea of what a seat includes. Comparing them is work before it's insight.
Nobody owns it. The split-mandate problem above. Not negligence. A gap in the org chart.
Isn't this just FinOps?
I thought so at first. I don't any more, and the distinction is worth being precise about.
FinOps is built around resources you provision and the tags you hang on them. Coding agent cost management is built around usage at the developer endpoint, which is agentless and scattered across vendors. The instinct is identical; the tooling is not. What people have always wanted from FinOps, knowing what they spend and spending it well, now has to reach a layer the existing tools were never designed to read.
It's the part of AI efficiency that finally reaches the keyboard. Complements, not rivals.
And this is the bit that convinced me I wasn't just inventing a category. On 4 August 2026 the Linux Foundation formally established the Tokenomics Foundation with 30 founding members, specifically to define the economics and ROI of AI. It works with the FinOps Foundation rather than replacing it. One of its first jobs is extending FOCUS, the open cost and usage specification, to account for token-based consumption.
PointFive joined as a Premier member and took a seat on the Governing Board. I'm obviously not neutral about that. But the relevant point isn't who joined, it's what the existence of the body tells you: when an industry forms a standards organisation for a cost category, that category has stopped being a rounding error. Alon put our reasoning as "standards first, tools second". Why AI cost needs a standards body is his argument, not mine.
What I think doing it well looks like
Five things. I originally had four here and left governance out, which in hindsight was the same mistake the industry keeps making.
- Visibility. One view of spend across every agent, in the same units, fresh enough to act on.
- Attribution. Cost tied to teams, projects, and ideally the work itself, so the number means something to someone who can change it.
- Governance. The ability to actually constrain it: budgets and limits per team, which models are approved for which work, and what happens when a threshold is crossed. Also the compliance half: whether code leaves your estate, which providers are sanctioned, where inference runs.
- Optimization. Real actions that bring spend in line with value: plan and model choices, and how people actually use the tools.
- Forecasting. An honest read on where this goes as adoption grows, so next quarter isn't a surprise.
Governance is the one I'd argue hardest for now, because the data says it's the gap. KPMG's 66% with dashboards against 36% with token or usage controls is a governance gap, not a visibility gap. And PointFive Labs found that 50-80% of a typical enterprise AI bill sits outside the in-platform spend controls teams rely on, running directly against model providers where no single console sees it. You cannot govern what your controls cannot reach.
Visibility is the floor, and it's where most people stop.
Where it sits
This is one layer of AI efficiency, not the whole story. It sits alongside managing production AI, data platforms, and cloud infrastructure. What makes it urgent is that it's the newest layer and the least covered, so it's where unmanaged spend piles up fastest. I wrote up the wider picture in the fourth layer.
If you're further along with this than I am, I'd genuinely like to hear where I've got it wrong.
FAQ
What is coding agent cost management? Measuring, attributing, and optimizing the spend that AI coding tools like Claude Code, Cursor, and Copilot generate across an engineering org.
Why isn't it just a software licence cost? Because coding agents are billed for consumption, usually by the token. Cursor, Codex and GitHub Copilot all moved to usage-based billing between June 2025 and June 2026. Two developers on identical seats can differ by an order of magnitude. DX recorded individual bills moving from $29 to $750 and from $50 to $3,000 a month.
What causes AI coding budget overruns? In DoiT's February 2026 survey of 500 finance leaders, accountability was split almost evenly: 55% said technology owns AI spend and 53% said finance does. No single function has a clear mandate, and 36% named lack of clear financial attribution as a core barrier. Overruns were also higher at the most FinOps-mature organisations (89%) than at early-stage ones (69%), which suggests the root cause is accountability rather than measurement.
Is it the same as FinOps? No. FinOps is built for provisioned cloud resources and tagging. Coding agent cost management deals with usage at the developer endpoint, which is agentless and spread across vendors, so it needs a different approach. They're complements.
Who should own it? It works best as a shared job between engineering, which picks and approves the tools, and finance, which carries the budget, with a platform underneath giving both sides the same view. The failure mode is assuming the other side has it.