AI software that drafts, reviews, and rewrites code is shifting from flat seat fees to usage based billing, and per developer bills metered by AI tokens are already climbing into the salary band of the engineers the tools sit alongside.
AI coding agents are accumulating a monthly invoice that, for some teams, already rivals the salary of the developer doing the work. These are the AI tools that draft, review, and rewrite code on behalf of engineers, packaged as chat assistants, IDE plug-ins, and autonomous background workers, and they now bill by the token rather than by the seat. According to a Gartner prediction reported by The Register, by 2028 the cost of running those agents will, on average, overtake the average developer's salary.
The shift is structural, not cyclical. The industry is moving from seat-based subscription pricing, a flat fee per engineer per month, to consumption-based pricing, where the bill scales with the tokens the agent consumes. In the seat model, doubling the number of developers doubled the bill, which gave procurement a clean per-headcount lever. In the consumption model, the bill scales with how much work the agent does, how many retries it runs, and how large a context window the team feeds it, none of which track cleanly to headcount.
Nitish Tyagi, a senior principal analyst at Gartner, told The Register that per-developer AI coding bills are jumping from a $20 to $100 a month band into a $2,000 to $5,000 a month band, with extreme cases pushing toward $20,000 a month in token charges. The crossover, in Tyagi's framing, is location-invariant: a developer's AI coding bill can already equal the salary of an engineer with four to six years of experience in India, even though US developer salaries are higher. The Register's write-up was independently corroborated by Computer Weekly, which carried the same 2028 prediction.
The structural risk is the incentive change underneath. In a consumption model, more tokens mean more revenue, and Tyagi described the prevailing vendor posture as "tokenmaxxing," marketing higher token consumption as productivity rather than engineering waste. None of the major AI coding vendors currently ship robust built-in cost optimization, in Gartner's read. The implication is that the cheapest way for a vendor to grow is for the customer to spend more, which inverts the procurement discipline that has governed developer tooling for the last decade.
That inversion is starting to show up in enterprise budgets. TechCrunch recently ran a piece on the industry's scramble to manage runaway AI costs, and a Forbes column from Janakiram MSV walked through how engineering favorites are wrecking 2026 IT budgets. Morph LLM's token-math comparison of Claude, Codex, and Gemini shows where the spend goes at the model level. The pattern across these reports is the same: redundant writes, oversized context windows, retry loops, and the human review overhead that follows each agent run.
Gartner's recommended mitigations are deliberately unglamorous: token consumption optimization, context engineering (tighter inputs and shorter prompts), and model routing, which sends the high-frequency, low-difficulty work to smaller, cheaper models and reserves frontier models for the hard problems. None of these are features a vendor will ship if the vendor's revenue model depends on selling more tokens.
What to watch next is whether the procurement side of the industry catches up. The Register's same-day companion piece on database vendors positioning around AI cost control is an early signal that the tooling market is starting to bifurcate: agent vendors on one side selling throughput, and the adjacent infrastructure vendors on the other side selling cost containment. Engineering leaders and their finance partners who treat agent inference as another workload-scaling expense, measured, gated, and routed by task, will absorb the crossover. Those who treat it as a flat seat will not see the bill coming until it is already past the developer's salary.