What is tokenmaxxing? (and what to do instead)

What does “tokenmaxxing” mean?

Tokenmaxxing joins “token”, the unit that language models are metered and billed in, with the “-maxxing” suffix people use for pushing one attribute as far as it will go. IBM’s Think explainer defines it as the organizational practice of incentivizing employees to maximize their AI token usage. In practice it looks like a leaderboard that ranks engineers by tokens used, a monthly token quota that teams are expected to hit, or an executive dashboard where “tokens per engineer” is read as an adoption score.

The word is not always an insult. Some leaders use it to mean “use AI aggressively so the organization learns faster.” That is not what this article is about. The problem is narrower. It is what happens when token volume itself becomes the target and people are rewarded for hitting it.

Our shorthand for the shift we recommend is tokenmaxx to token efficient. Keep the ambition to use AI heavily. Change what gets measured, from tokens consumed to outcomes shipped per dollar.

Why did tokenmaxxing catch on, and why did the narrative flip?

In late 2025 and early 2026, token consumption was the easiest evidence that AI adoption was happening. Frontier coding agents made it easy to spend. Per-token prices kept falling. Leadership wanted proof that expensive seats were being used, and a token leaderboard is cheap to build and easy to read. By April 2026, Business Insider was covering the debate after reports that engineers at large tech companies were competing on internal token dashboards. Some investors defended the practice. Others called it reckless.

The mood changed within weeks. Faros AI called tokenmaxxing the AI-era version of measuring developers by lines of code, and argued that token consumption is an input, not an outcome. Lance Eliot wrote in Forbes that companies with tokenmaxxing goals were foolishly inspiring employees to waste costly AI resources. On May 28, 2026, Fortune declared tokenmaxxing dead, because companies did not get the ROI from AI they wanted to see. IBM’s explainer, published in September 2026, describes the practice as peaking in spring 2026 and fading since.

None of this should have been surprising. Goodhart’s law says that when a measure becomes a target, it stops being a good measure. Token count was a reasonable proxy for adoption right up until people were rewarded for it. Then it got optimized directly. Larger contexts, always-frontier models, agents left running in loops, failed tasks re-run for volume. IBM’s explainer points to reporting of staff inflating token spend on irrelevant tasks just to hit internal usage targets.

Tokenmaxxing vs token efficiency

What should engineering leaders do instead?

Here are the five steps we recommend, in the order we would run them.

1. Replace token leaderboards with outcome metrics. Count merged pull requests, production deliveries, review coverage, or Mentlio Delivery Points, which is our normalized unit of shipped engineering work. Rank workflows, not people. Our guide to measuring AI ROI in software engineering walks through the outcome units.
2. Attribute spend before you try to change it. Know which tool, model tier, team, and workflow the bill comes from. The vendor invoice gives you a total. Metering on the developer’s device gives you the shape of it.
3. Compute cost per shipped outcome. Divide attributed AI spend by outcome units, per team, per period. Then watch the trend. The direction matters more than the absolute number.
4. Remove waste at the agent loop instead of capping usage. Most of a coding task’s cost is created before the model is ever called. Redundant file reads, repeated tool output, oversized logs, and frontier models on routine work. Mentlio’s Token Savers (Route, Lens, Quiet, Logs, Recall) go after exactly that. Route picks the cheapest capable model. Lens finds the right code before the agent opens files. Quiet shortens agent output without dropping paths or warnings. Logs collapses repetitive tool output while keeping every error. Recall keeps large outputs on the device until the agent needs them. Our product thesis is that routing, context optimization, and prompt compression can remove on the order of 30% of AI coding spend without capping useful usage. That is a thesis about where the waste lives. It is not a customer result.
5. Keep the measurement private. Derived metrics are enough to run this program. Token counts, cost metadata, classifications, and scores. Raw prompts, model outputs, and source code should stay on the developer’s machine.

How Mentlio connects cost to accuracy

The usual objection to token efficiency is that cheaper means worse. That is true of naive cost-cutting, and it is why “cap the budget” is the wrong move. Our approach is to hold a quality floor and then find the lowest cost above it. We describe how in A more useful way to manage AI coding costs.

In our frozen Terminal-Bench 2.1 replay, we saw Mentlio Route retaining 98.93% of Fable 5’s solved-task rate at 24.90% lower average per-task model cost. Route solved 82.92% of tasks at $0.933 per task against Fable 5’s 83.82% at $1.242 per task. Those numbers come from a replay over a fixed set of 89 tasks, not from a production deployment, and the write-up states its assumptions. Read them as evidence that the tradeoff can be managed, not as a promise about your workload.

Sources: IBM Think, “What is tokenmaxxing?”; Business Insider, April 2026; Faros AI, “Tokenmaxxing”; Forbes, May 19, 2026; Fortune, May 28, 2026. Our benchmark figures are documented on the research pages. They are replay results, not customer case studies.