Gartner says your AI coding agents could cost more than your developers — by 2028

One developer burned $4,000 in three days on an AI coding agent — not from misuse, just from normal agentic coding sessions left running. Gartner now projects that at current growth rates, AI coding agent spend could exceed the average developer's salary by 2028. Most engineering budgets still treat it as a seat license, not a consumption bill.

The seat fee was never the real number

Coding assistants got budgeted the way most SaaS tools do: a monthly per-seat license. GitHub Copilot Enterprise and Cursor Business both land in the $40-60 per seat per month range — a predictable line finance can plan around. But agentic coding tools bill token consumption on top of that seat fee, and that number moves independently of headcount. Combined real-world cost, seat plus tokens, runs $200-600 per developer per month — $2,400-$7,200 per year — before counting the outliers.

The outliers are the budget risk

The averages understate the exposure. One developer accidentally spent $4,000 in three days running Claude Code sessions. Individual heavy users on support teams have reached $800 a month on their own. Nearly 25% of tech leaders already spend $200-500 per developer per month on AI coding tokens, and about 6% spend over $2,000 per developer per month. One 80-developer organization found its combined monthly AI coding spend equaled a full-time engineer's annual salary.

Why Gartner's 2028 projection matters now

Gartner's research attributes the coding-agent-versus-salary crossover to "the rapid growth of LLM token consumption" — meaning this isn't purely a pricing story, it's a usage-growth story. The more autonomous and multi-step coding agents get, the more tokens a single task burns, and adoption is still accelerating, not leveling off. A 100-developer organization already reaches $400,000-$600,000 a year in combined seat and token spend — a number few engineering budgets were built to absorb when the tool was first approved as a seat license.

Who actually owns this budget line

Engineering managers typically approve coding-assistant rollouts as a productivity investment, without FinOps involvement, because the request started life as a seat license — the same approval path as any other developer tool. The token-consumption side of the bill inherits none of the governance built for the rest of the AI budget — model routing, anomaly alerts, per-feature attribution — unless someone explicitly extends those controls to cover it. Most organizations haven't, because the coding-tool line still doesn't read as "AI spend" in the same budget review that scrutinizes the product's API costs.

What to check before the next seat expansion

• Whether coding-tool spend is tracked separately from other LLM usage, or sits invisible inside a single vendor invoice.
• Per-developer cost distribution, not just the average — the outliers are where the $4,000 weekends happen, and an average hides them completely.
• Whether an anomaly on a single account gets flagged the day it happens, or only shows up the month the invoice arrives.

The bottom line

A seat license was never going to be the real number for agentic coding tools — token consumption was always the variable that could grow past it, and by 2028, Gartner's projection says it might grow past the developer's salary too. The fix isn't slower adoption. It's treating coding-agent spend as what it already is: a consumption-based AI cost line that needs the same visibility as every other model call in the stack, not a seat count on a SaaS renewal.