Per-seat pricing is quietly ending

For a decade, software costs were predictable. You paid a fixed fee per user per month, and the bill barely moved no matter how hard anyone worked. That model is now breaking for AI tools. In June 2026, a leading AI coding assistant moved its plans to usage-based billing, charging for the tokens each request consumes rather than a flat seat price.

The shift matters because AI agents do not behave like a person clicking through a menu. An autonomous agent that plans, edits, and checks its own work can consume on the order of a thousand times more tokens than a single question. The same seat that cost a fixed amount last year now bills against an open meter, and the heaviest users run it hardest.

The numbers are already alarming

The early reports are not subtle. Heavy agentic users have seen projected costs rise tenfold to fiftyfold under the new metered model. One large technology company is reported to have spent its entire annual budget for AI coding tools within four months, then capped each employee at a fixed monthly amount per tool. Even one of the largest software companies in the world has been reported to cancel most of its internal licenses for an outside AI tool at the end of June, citing runaway token costs.

There is a deeper trap. Token prices have fallen sharply since 2023, yet total AI bills have risen, because cheaper tokens invite far heavier use. Lower unit prices do not protect a budget when consumption is uncapped. The cost line that owners were told would shrink with scale is the one growing fastest.

The credits expire at the end of summer

Much of the current pain is being softened by promotional credits that vendors attached to the billing change. Those credits are temporary. In the most prominent case, the cushion runs only through the summer of 2026 and then stops, leaving the full metered bill in place from autumn onward.

This makes the timing concrete. An organisation that has not measured its real, uncapped AI consumption before the credits lapse will discover the true number on an invoice rather than in a plan. The quiet months are the window to install controls, not the moment after the meter is fully exposed.

Governance, not a cheaper tool

The instinct to switch to whichever model is cheapest this quarter misses the point. The real exposure is the absence of attribution and limits. Most organisations cannot say which team, which agent, or which workflow drove a spike, which means they cannot govern it. Industry surveys now show almost every company actively trying to manage AI spend, while readiness to actually govern it lags well behind.

The owner action is straightforward and not technical. Demand per-team and per-agent cost attribution, set hard caps before the promotional credits expire, and make the choice of model a deliberate decision for each job rather than a default. AI is now a metered utility. It needs to be run like one.

Read next: Match the Model Tier to the Job  ·  When You Cannot Buy The Best Model