After tokenmaxxing: why token spend is a cost, not a productivity metric

The logic was simple. If AI makes engineers faster, an engineer who uses more AI must be getting more done. So usage became a signal of ambition, and the signal turned into a scoreboard. Reported examples from 2026 show how far it went:

• Leaderboards. The New York Times reported internal rankings of token usage at large AI-heavy companies, and Fortune reported that one such leaderboard at Meta covered tens of thousands of employees and tens of trillions of tokens a month.
• Tokens as compensation. Nvidia’s Jensen Huang floated giving engineers a token budget worth around half of their base salary, and Databricks’ Ali Ghodsi said he would be alarmed if a highly paid engineer was not consuming a large share of their salary in tokens.
• Salary-sized bills. Individual engineers told reporters they likely spend more on AI than they earn, with the employer covering the difference.

None of these people were being irrational. Early in a technology shift, heavy use really is a reasonable proxy for learning. The problem starts when the proxy becomes the goal.

By spring the stories changed tone. Uber was reported to have used up its entire 2026 AI budget within about four months, and its leadership said publicly that the link between AI spend and what reached customers was not yet clear. Analysts began forecasting that AI coding costs per developer could approach developer salaries within a few years, largely because licensing is moving from seats to consumption.

Telemetry from engineering-analytics vendors added a second, less comfortable picture. Teams with heavy AI adoption completed more tasks and merged more pull requests, but they also reported more bugs per developer, longer review times and more changes merged without review. More output went in; not all of it came out the other side as value.

The spend was easy to see. What it bought was not.

Measuring engineers by tokens repeats the oldest mistake in software metrics: counting lines of code. A large number can mean a lot of useful work, or a lot of motion. With coding agents the second case is common, for a structural reason.

An agent session is a loop. Every turn sends the growing context back to the model, so input tokens, not the code the agent writes, make up most of the bill. Published breakdowns of agent sessions show input outnumbering output by more than twenty to one. A session that keeps retrying, re-reading files and drifting from the task burns tokens faster than one that finishes cleanly, and two engineers using the same tool on similar work can differ by an order of magnitude in cost.

A leaderboard cannot tell these rows apart. It rewards all of them equally.

The natural reaction to an exhausted budget is a cap. Per-seat limits, a forced switch to cheaper models, a freeze on premium plans. That stops the bleeding, but it uses the same broken metric in reverse: it treats every token as waste instead of every token as progress.

• A flat cap punishes the engineers whose expensive sessions actually ship, along with the ones that thrash.
• A cheaper model can cost more per finished change if it needs more turns and more retries to get there.
• A freeze hides the problem until the next budget cycle, without telling anyone which workflows were worth the money.

Token spend is a real number and it belongs on the dashboard. It just belongs next to the work it paid for, not on its own. The shift is from “how much did we use” to “what did each shipped change cost, and what cost money without shipping”:

1. 1

   Cost per shipped change

   Follow a change from task to AI session, commit, pull request, CI and deploy, and add up what it cost along the way.
2. 2

   Spend that ended nowhere

   Sessions without a commit, pull requests closed unmerged and failed CI runs are where tokenmaxxing quietly turns into waste.
3. 3

   The basis of every number

   Know which figures were reported by a provider, which were measured and priced, and which are estimates. A missing price is a gap, not a zero.
4. 4

   Projects and workflows, not people

   The useful question is which kinds of work AI makes cheaper. Ranking individuals by spend, in either direction, recreates the leaderboard problem.

This is the view Nythrex is built around: the cost of every change from AI prompt to production, with the basis of each figure shown next to it, and named rules that point at money that produced no result. It does not rank people or rate productivity. How the chain works · What the recommendations look like.

Tokenmaxxing was a phase, and probably a useful one: it pushed teams to actually learn the tools. The next phase is less exciting and more important. AI spend is turning into one of the largest lines in the engineering budget, and like every large line it will be asked to explain itself. Teams that can show what each change cost, and where money went without shipping anything, will keep their budgets. Teams that can only show a token count will have them cut.

• Faros AI — Tokenmaxxing: why token consumption isn’t AI engineering productivity (2026)
• Forbes — After tokenmaxxing, token spend has become the new metric to watch (July 2026)
• CNBC — Nvidia’s Huang pitches AI tokens on top of salary (March 2026)
• Forbes — The ‘AI gods’ spending as much as they can on AI tokens (March 2026)
• Vantage — The hidden cost driver in agentic coding sessions (April 2026)