A new line item is showing up on corporate earnings calls and in C-suite spreadsheets: the AI token bill. Wired reports that companies from the communications firm 8x8 to Cisco are tallying fast-rising spending on AI usage, with some celebrating the productivity and others quietly capping how much their teams can consume. The awkward part is the shape of the cost. With enterprise tools like Claude Code priced as a seat fee plus actual token usage, spend scales directly with how useful the tool is: the better the agent and the more it does, the bigger the bill.
That creates a genuinely strange incentive, where success and cost rise together. The more capable agents become, and the research suggests the most expert users now trigger roughly a dozen model actions per prompt, the more tokens they burn, and the line between a productivity win and a runaway invoice gets thin. The anecdotes are already vivid: one developer reported running up nearly $31,000 in token usage in a single month on a nominally $200 plan. For a finance team trying to forecast next quarter, 'it scales with how useful it is' is not a reassuring sentence.
This is the cost side of the same story playing out across the agentic shift, and it connects directly to why pricing keeps getting reworked, including Anthropic's own paused attempt to meter its Agent SDK separately. The unresolved question is who absorbs the variability: the vendor, through flat-rate plans that risk losing money on power users, or the customer, through usage bills that punish exactly the heaviest and most productive use. The caveat is that token prices have fallen steadily and efficiency keeps improving, so today's sticker shock may not be permanent. But for now, tokenomics is a real constraint that companies are pricing into their AI bets.