The bill for artificial intelligence has stopped behaving the way software bills used to behave. For years, enterprise technology costs were predictable: a per-seat license, a flat monthly fee, a number finance could plan around. That predictability is disappearing fast for AI. As major providers shift from flat-rate subscriptions to consumption-based pricing, and as agentic systems multiply the number of model calls required to finish a single task, token spend has become one of the most volatile line items on the corporate ledger. The organizations pulling ahead are not necessarily the ones spending the most on AI. They are the ones that have learned to treat token supply as a resource to be managed with the same discipline applied to capital, headcount, or compute.
THE SHIFT FROM FLAT-RATE TO CONSUMPTION PRICING
The clearest signal of this new era is the pricing shift underway across the industry. GitHub’s move away from fixed per-seat licensing toward pure usage-based pricing marked a turning point for how enterprises pay for AI coding tools, and other vendors have followed with steep increases of their own, including model cost multipliers that pushed some legacy plans up by as much as 9x to 18x in a single update. The comfortable predictability of a flat monthly developer seat is giving way to variable consumption that can swing wildly month to month.
The scale of that swing has already produced some startling headlines. One company reportedly burned through $500 million in tokens in a single month after failing to set employee usage limits, while other major enterprises have blown past their entire annual AI budgets within the first few months of the year. These are not isolated accidents. They are the predictable result of two forces compounding at once: falling per-token prices paired with a sharp rise in the sheer quantity of tokens consumed per task, since multi-agent systems can use roughly 15 times more tokens than a simple chat interaction, and agentic coding workflows can run into the thousands of times more.
WHY TOKEN BUDGETS ARE HARDER TO PREDICT THAN THEY LOOK
Part of what makes token supply so difficult to manage is that AI systems are notoriously bad at forecasting their own consumption. Research on frontier models has found they systematically underestimate the number of tokens a task will actually require, largely because agent trajectories are inherently probabilistic. An agent does not know in advance how much context it will accumulate, how many steps it will take, or how many times it will loop back to revise its own output. Running the identical agent on the identical task can produce cost swings of up to 30 times between runs.
This unpredictability is compounded by where the money actually goes. Surface-level dashboards showing total monthly spend rarely reveal the underlying drivers. Every tool call an agent makes inside a Model Context Protocol integration adds tokens to the context window, and it does so regardless of whether the retrieved information was ever used in the final output. Redundant tool responses get reprocessed in full even when the underlying data has barely changed since the last request. Default configurations often fire tool calls on every single interaction, checking permissions or loading templates that are rarely relevant, quietly inflating every request with overhead nobody asked for.
MANAGING TOKENS LIKE CAPITAL
The organizations getting this right have started applying a simple reframe: tokens are not just a cost to minimize, they are capital to be deployed toward the highest-return opportunities. Just as a firm with capital to invest asks where that money will generate the best return rather than simply how to spend the least, leading AI programs are learning to direct token spending toward the workflows where the payoff compounds over time.
That reframe has produced a new class of metric known as return on intelligence, which divides the value of an AI-driven output by the combined cost of the labor and tokens used to produce it. The advantage of a metric like this is that it lets a company compare very different applications of AI on the same basis, whether that’s a customer service system substituting for human labor or a research process augmenting engineers with faster iteration cycles. Companies that measure only labor savings tend to underinvest in AI’s broader potential, while companies that optimize purely for maximizing token consumption create incentives for employees to game the numbers rather than create real value.
The payoff for getting this right shows up in the data. An analysis of more than a hundred public technology companies found that firms in the highest quintile of token usage posted median year-over-year revenue growth of 16.5 percent, compared with just 5.1 percent for the lowest-usage group. Consumer goods company Reckitt offers a concrete example of what that looks like in practice: by applying AI selectively across marketing, product development, and R&D rather than everywhere at once, the company reported up to 60 percent faster content development and faster research cycles with fewer prototypes.
WHERE THE WASTE HIDES
Industry analysis suggests that somewhere between 50 and 80 percent of typical token spend is unnecessary, and the waste tends to cluster in a handful of predictable places. Input tokens, not output tokens, drive the larger share of cost in most agentic workloads, which means system prompts sent on every single invocation, bloated file formats like screenshots and verbose JSON, and poorly implemented web search that floods a model’s context with duplicate snippets and navigation clutter are often bigger cost drivers than the actual reasoning work.
Verification and review carry a surprising share of the burden as well. One analysis of software engineering tasks found that the iterative code review stage accounted for nearly 60 percent of total token consumption, meaning the primary cost of agentic coding isn’t the initial generation of code but the repeated cycles of refinement that follow it. That has a human cost layered on top of the token cost too. As AI accelerates code generation, some organizations have reported a sharp rise in bugs per developer and a corresponding jump in code review time, which means the “fully loaded” cost of a token includes the hours engineers spend catching and fixing what the model got wrong.
Multi-agent systems introduce their own version of this problem. When agents review each other’s work in a loop, without a firm stop condition, a minor disagreement over formatting can spiral through five or ten rounds of revision, with each round dragging the full conversation history along and burning tens of thousands of tokens on what started as a trivial edit.
THE MULTI-MODEL PLAYBOOK
The single most effective lever available to most organizations is simply refusing to send every task to the most expensive model available. Open-weight models have closed much of the performance gap with frontier systems, and for a large share of coding requests, a mid-tier model can match frontier performance at a fraction of the cost. Smaller specialist models can also do preliminary work, searching, sorting, and reranking information, before handing a distilled summary to a more expensive model, cutting token use substantially while preserving accuracy.
This has given rise to a growing category of AI routers and gateways, tools that select the most appropriate model for each request based on cost, latency, and quality requirements, then layer on governance features like authentication, budget controls, and observability. The logic mirrors how a manager might assign work inside a company: a routine task doesn’t need the most senior, most expensive person on the team, and a model request doesn’t always need the most capable, most expensive model on the market.
There is a strategic dimension to this diversification beyond cost. Reliance on a single frontier provider carries its own risk, since access to any given model can be disrupted by regulatory action or export controls with little warning. Building systems around multiple models and providers, including open-weight options an enterprise can run and control itself, has become as much a resilience strategy as a cost strategy.
GOVERNANCE, CAPS, AND POOLED BUDGETS
As the financial exposure has grown clearer, a number of major enterprises, including Uber, Coinbase, Match Group, and Walmart, have introduced caps on how much individual employees can spend on AI tools. At the same time, most organizations are wary of setting rigid budgets too early, since usage patterns are still shifting rapidly and heavy-handed limits risk discouraging the experimentation that produces the biggest wins. A common pattern that has emerged is the “pooled spend” model, in which a small share of power users, often around 5 percent of employees, account for a disproportionate share of total consumption, and that pool shifts in size and composition as adoption matures.
Visibility is the precondition for any of this working. Effective token governance requires answering specific questions: which teams are generating the highest costs, which models are being called and at what price, which projects carry the heaviest consumption, and whether spending is actually correlated with productivity gains rather than inefficient or redundant workflows. Without that level of granularity, a sudden CFO inquiry into a budget overrun tends to produce a blanket cutback that penalizes efficient use alongside wasteful use in equal measure.
THE OUTLOOK
Token supply is unlikely to become more predictable anytime soon. Prices per token will keep falling, agentic workflows will keep multiplying the number of calls required to complete a task, and the gap between commodity and frontier inference will likely keep widening. What separates the organizations winning this new form of competition from those simply absorbing rising bills is the discipline to treat tokens the way any other scarce, valuable resource is treated: measured carefully, routed intelligently, and directed toward the work that generates the greatest return.
References
Kropp, M., Bedard, J., O’Niell, C., Duranton, S., and Hsu, M. “The Era of Token-Based Competition Is Here. Is Your AI Strategy Ready?” Boston Consulting Group, June 18, 2026. https://www.bcg.com/publications/2026/the-era-of-token-based-competition-is-here
“How to Manage AI Token Costs Before They Become Your Largest IT Expense.” Airia, July 6, 2026. https://airia.com/blog/how-to-manage-ai-token-costs-before-they-become-your-largest-it-expense/
Jalan, A., Menashy, S., and Roy, P. “AI Tokenomics: How to Tokenmin While ROImaxxing.” MMC Ventures, June 30, 2026. https://mmc.vc/research/ai-tokenomics-how-to-tokenmin-while-roimaxxing/
