Tokens Are the New Lines of Code

Back in my programming intro course at university, we had to build a basic text editor in Pascal. We had nailed every feature, but minutes before evaluation someone noticed we forgot one requirement: the program must have at least 1000 lines. We had implemented everything in about 600. Pascal happens to have a lovely property: once the compiler finds end., it silently ignores the rest of the file. Guided by holy laziness, I didn't hesitate: I copy-pasted a bunch of blocks past end. and doubled the line count. We got 100%.

The same professor used to say that a good software engineer ships only a few lines of code a day. But as an evaluator, she couldn't find a better metric to estimate a student's ability.

Fast forward. Big companies now measure how many tokens developers consume on the job. From a software engineering standpoint, it's so insane I don't know where to start. Lines of code was already a terrible metric: it rewards verbosity, punishes concision, and says nothing about correctness or impact. Measuring token usage takes the same problem and injects it into the AI era. It's the same problem: you optimize what's measured, not what matters. Managers get a dashboard, developers get perverse incentives, and everyone loses.

But I'm not writing this just to blame managerial laziness. The pattern is predictable: you have a complex, qualitative activity; you need a number; you pick an easy one. LoC was easy. Tokens consumed is even easier (the AI vendor provides it). Neither is good, but tokens is fundamentally worse.

Here's my small proposal. Let's put LoC and token usage together in the same metric. One in the numerator, the other in the denominator.

tokens consumed lines of code written

The logic is simple: coding assistants perform better in small, focused codebases. Keeping code tidy has always been important, but now it's critical: AI tools produce verbose, unmaintainable code by default, and a tangled codebase makes them worse. A high tokens/LoC score means you're extracting real value from the assistant without letting it bloat your code. It pushes you toward concision, toward cleaning up as you go, toward the kind of hygiene that makes AI-assisted development actually sustainable.

Is it a good metric? No. But it's a small improvement over measuring raw token consumption, and a better conversation starter than a dashboard that says "developer consumed 42k tokens yesterday."