Spend fewer tokens, spend more — both companies are half right

2026-06-26 · Ascendy Engineering

TL;DR

Two opposite cultures exist. Token efficiency (spend fewer tokens = competent) and tokenmaxxing (spend more = competent). One rewards saving tokens in reviews; the other treated token burn as a measure of innovation.
They look opposite, but both measure token volume. Same error, reversed sign. It’s the AI era’s overtime hours — the token version of “whoever sits longest works hardest.”
I do the same work but spend tokens freely on my personal account and sparingly on my company one. It looks contradictory, but there’s a reason — the billing structure and the bottleneck structure differ.
The real axis isn’t tokens. It’s whether the time those tokens buy sits on the organization’s bottleneck. So the right evaluation isn’t measuring tokens directly — it’s setting a generous cap, and inside it judging productivity, not token count.

About this piece. This column came out of an interview (/interview) — the operator answered, an AI editor pressed, and the position developed. Assertion and speculation are kept apart (the bottleneck logic and the prescription especially are one operator’s N=1 observation). No company name is used; public phenomena are cited as such.

Two camps — efficiency and tokenmaxxing

As organizations started folding AI usage into evaluation, two opposite cultures emerged.

One is token efficiency. Doing more with fewer tokens is read as competence. One large company reportedly measures “how many people’s worth of work one AI-using employee does” as a full-time-equivalent figure, and added output per token spent as a secondary metric on top. More broadly, token usage itself is increasingly treated as a performance yardstick.

The other is tokenmaxxing. In Silicon Valley, token consumption was for a while treated as a measure of innovation capacity, with internal leaderboards built to encourage spending more. The result was burning through budgets in a few months — waste — and it’s now being walked back.

On the surface they’re opposites. But step back and both use token volume as a proxy for performance. One says less is better, the other more, but they both stare at the token count.

This is the AI era’s overtime hours

Doesn’t this setup look familiar? It’s the “whoever works the most overtime works hardest” illusion.

Overtime hours are easy to measure. So they once served as a proxy for diligence. But we know now — doing less and producing more is best; sitting longer isn’t ability. Token usage is the same. It became a metric only because it’s easy to measure.

More honestly: nobody yet knows the real standard for AI productivity. With no standard, people cling to the number they can hold (tokens) — a transitional fumble. Just like overtime.

Why the same person acts in opposite ways

Building a product solo, I spend tokens without hesitation on my personal account. On the company account, I save them. There are two reasons the same person behaves in opposite ways.

First is the billing structure. A personal subscription is flat, so you can spend fairly freely (honestly, today’s subscription models appear — my speculation — to run at a loss; the user is being subsidized), while an enterprise plan bills in proportion to usage.

The second is more fundamental — the bottleneck structure. In a large organization, if one employee runs short on tokens and finishes a little late, it has essentially no effect on total revenue. And in a research-leaning function, a slipped schedule doesn’t break the company. That one person isn’t on the critical path.

My solo project is the opposite: my speed is the organization’s speed. Because it’s just me. If I slow down, the whole project slows, and at worst the output never ships. I’m a single point of throughput.

So the real value of a token isn’t in the token count. It’s in how directly the time it buys sits on the organization’s bottleneck. Token efficiency and tokenmaxxing aren’t wrong — each is a locally rational response to its own bottleneck structure. What’s wrong is forcing one axis (token volume) on everyone alike.

A cheap check wasn’t a check

The easiest way to save tokens, honestly, is to hand the work to a cheaper, lighter model. So I tried slotting in the Gemini family — what I considered the best value — as an adversarial code reviewer. I gave it the exact same instructions (AGENTS.md) I gave other reviewers.

The limit was clear. When I had Codex re-review a PR that Gemini had “approved,” real feedback came back. And even the Claude that wrote the code conceded it was a flaw it had missed, and fixed it. Same instructions — the model’s weight class decided the outcome.

This is the point. I used a cheap reviewer to save tokens, and saved barely any — while losing the most expensive thing, my time. The cost of re-running a wrongly-passed review and catching the missed flaw later far exceeded the tokens I’d saved. A cheap check wasn’t a check.

That doesn’t make efficiency wrong

Don’t misread me. More output with fewer tokens is clearly best — obviously so, for whoever pays the bill. The aim of token efficiency is right.

There’s a reason companies leaned into tokenmaxxing too. Push “save tokens” too hard and you get the opposite failure: people who could deliver instead hold back, or use tokens as an excuse not to. Tokenmaxxing was a reaction to that flinch. So both camps hold a partial truth and carry a side effect.

The trouble starts the moment you nail an aim directly onto an evaluation metric. Start measuring tokens directly and people optimize the metric instead of the work. Tell them to spend less and they flinch; tell them to spend more and they waste.

So what’s the right direction

Here’s my answer. Set a generous cap; inside it, let people spend tokens freely — and make the thing you evaluate productivity, not token count.

That solves two things at once. The cap controls cost, so tokenmaxxing-style runaway is blocked. And inside it, people focus purely on results without second-guessing whether to spend more or less. Then the optimal token spend isn’t set by an evaluator — each person finds it as a result.

It’s the same principle as keeping regulation minimal and letting the market find the right price. The right token spend isn’t a number to be fixed from above — it’s an equilibrium that emerges from working freely inside a constraint.

When there’s no measurement standard yet, the move isn’t to force a fake metric on everyone. It’s to set the constraint and judge by results. Just as we moved past the era of rating people by overtime hours, this era of rating them by token usage is a phase we’ll move past too.

Authorship & citation: Written by Ascendy Engineering; quotable with attribution. Found something wrong? Let us know via a GitHub issue.

Tags: ai, productivity, token-economy, measurement, opinion