From Tokenmaxxing to Token FinOps:The Agentic AI Bill Comes Due

A full-year budget gone in six months

Over the past few months, an unusual number of companies have run into the same problem: by midyear, their forecast annual token budgets were already fully spent, and in some cases far exceeded. Among the companies that have gone public about it, Uber said it had used up its entire 2026 budget for AI coding tools in four months, and its COO has publicly acknowledged that the link between that spending and the features delivered to customers “is not there yet.”

The cheap-token paradox

Token prices are plunging. According to a 2025 Epoch AI study, the price of reaching a given level of performance falls by a factor of 9 to 900 every year, depending on the task. A recent lightweight model often matches the flagship model of a year or two ago at a much lower cost.

But the total bill keeps rising. Teams are quick to adopt the latest models, whose prices fall much more slowly. Each task also uses more tokens, because reasoning models generate thousands of invisible “thinking” tokens, agents chain dozens of calls for a single request, and context windows keep getting longer. Finally, a use case that works gets rolled out to other teams, which multiplies its consumption.

This is the Jevons paradox, which holds that when a resource becomes cheaper and more useful, its total consumption goes up. The increase usually means use cases are working, but it still needs to be kept under control.

Tokenmaxxing, Tokenomics, and Token FinOps

Tokenmaxxing is the race to consume tokens, fueled by internal leaderboards that measure AI adoption by the volume each employee or team consumes. At Meta, one such leaderboard logged more than 60 trillion tokens in thirty days before it was shut down once the press picked up the story. Uber, for its part, ranked its teams by usage, which accelerated consumption. Employees end up burning tokens with no business purpose, simply to climb the rankings.

Tokenomics covers the economics of the token, meaning the way spending converts into tokens, then into value.

Token FinOps is its operational counterpart: the set of practices used to measure that spending and act on it. According to the FinOps Foundation, 98% of FinOps teams now manage AI spending, up from 31% two years earlier.

Seeing and acting: two underestimated challenges

The first obstacle is visibility. Tokens are consumed across several model providers and through many channels: direct API calls, hyperscaler platforms, coding assistants, AI embedded in SaaS applications, not to mention shadow AI. Each channel has its own billing unit (tokens, licenses, credits, volume commitments). Piecing together who consumes what, for which use case and with which model, is a project in itself. Most of the companies that reached out to us knew their total bill, but rarely its breakdown.

The second is the ability to act. Capping usage or banning certain models first requires the technical means to do so. Without a single entry point for model calls, with API keys scattered across the organization and subscriptions purchased directly by individual teams, there is no way to enforce such rules. The choice is also strategically delicate, because companies are trying to get dozens of use cases to work at the same time. Throttling consumption too early would hold back the most promising ones, while having no rules at all leads to the overruns we saw this year.

From the most powerful model to the right-sized one

Companies still do little to optimize model selection. Most teams default to the most powerful model available. That choice makes sense in the pilot phase, when speed and reliability come first. It is harder to justify at scale, because a large share of requests (rephrasing, classifying, extracting, summarizing) does not need the best model on the market.

Smart routing is one way to address this problem. Instead of sending every request to the same model, a router assesses each one before dispatching it. In practice, simple heuristics are enough to sort requests into three tiers: standard, complex, or requiring in-depth reasoning. They draw on the length of the prompt and context, the type of task, the presence of code or calculations, the number of steps expected, or how critical the use case is. Each request goes to the model that matches its tier, with automatic escalation to the next tier up if the response fails quality checks. Unlike a gateway, which centralizes calls and enforces caps, a router picks the model request by request. OpenAI already applies this principle in GPT-5, which includes a router that chooses in real time between a fast response and a reasoning mode. In research by LMSYS, routers cut costs by 35% to more than 85% depending on the benchmark, while retaining 95% of the top-tier model’s quality.

For a team to accept a cheaper model, it needs proof that the model is “good enough.” A/B testing provides that proof by routing part of a use case’s traffic to the cheaper model, then comparing quality (automated evaluation, user feedback, human rework rate) and cost per successful task. That metric is more accurate than cost per token, because an expensive model that gets it right the first time can cost less than a lightweight model that needs three attempts and human rework. Since the model landscape keeps shifting, the test has to be rerun, and the routing adjusted, with every new generation of models.

Five workstreams for Token FinOps

Several categories of solutions are taking shape, including AI gateways and routers, specialized observability platforms, FinOps software extended to AI, and application monitoring tools that now track consumption down to the token. The market remains fragmented, and no tool can replace a proper operating model. We structure our approach around five workstreams.

1. Enterprise-wide monitoring of AI usage. Consolidate all spending, across every provider and channel, into a common taxonomy: use case, business unit, team, model. The other workstreams depend on it.
2. Governance and allocation models. Assign every dollar to an owner, charge costs back to business units (or at least show each unit what it consumes), replace the fixed annual budget with per-use-case budgets revised as adoption evolves, and adjust supplier contracts accordingly.
3. Agent observability. Monitoring tells you how much; observability tells you why. Tracing each agent run reveals loops, retried calls, and overloaded contexts, and shows what a typical run costs.
4. Control mechanisms and guardrails. Budgets per agent or per team, alerts, soft limits before hard ones, circuit breakers for runaway agents, and model access rules by type of use rather than blanket bans. These rules work best when they vary with the maturity of the use case: a looser budget during exploration, a stricter framework once the use case is in production.
5. Architecture and model optimization. A router rather than a simple gateway, A/B tests rerun with each new model generation, caching, batch processing, and context hygiene.

Managing for value

As Uber’s COO put it, if you cannot draw a direct line between the spending and the features you ship to users, “that trade becomes harder to justify.” Token FinOps is about knowing what each token delivers, which may mean consuming more where the value is proven and cutting back elsewhere.

A few questions can help a leadership team take stock:

— Do we know the cost and the value of our top ten AI use cases?

— Do our model calls go through a single entry point that can control and route them?

— Who has the authority to trade off a more expensive model against a better outcome?

Even at the most advanced companies, the Tokenmaxxing of the first half of the year has given way to Token FinOps. Over time, organizations will need to tie their AI spending to the value it creates.

Sources and recommended reading

1. Fortune, “Uber burned through its entire 2026 AI budget in four months. Now its COO is questioning whether it’s worth it,” May 26, 2026.
2. Epoch AI, “LLM inference prices have fallen rapidly but unequally across tasks,” March 12, 2025.
3. The Pragmatic Engineer, “The Pulse: “Tokenmaxxing” as a weird new trend,” 2026.
4. FinOps Foundation, State of FinOps 2026, February 19, 2026 (1,192 respondents).
5. OpenAI, “Introducing GPT-5,” August 2025.
6. LMSYS, “RouteLLM: An Open-Source Framework for Cost-Effective LLM Routing,” July 1, 2024. Results obtained on public benchmarks with 2024 models: an order of magnitude, not a promise of enterprise gains.