Who would have thought we could find an even worse measure of developer productivity than lines of code?
It seems that the software developers at Meta, who are all in on AI-powered coding, came up with a notion they called “Claudeonomics” to measure their all in-ness. This manifested itself as an internal dashboard/scoreboard of who was burning the most tokens with Claude Code. The race was on to see who could burn through the most tokens.
Never mind whether this conflagration of Claude tokens was actually producing anything good. The chart merely gave boasting rights to the developer who cranked through the most processing power, declaring leaders as “Token Legend” and “Cache Wizard.”
Similar things were going on at Microsoft and Salesforce.
This is just the latest chapter in an age-old battle, and it is a really bad idea.
Maximum bad
Managing software developers is hard enough. There are many reasons why managing developers is hard, but chief among them is that it is difficult (if not impossible) to measure the process of writing software.
And that isn’t from a lack of trying. We’ve measured lines of code, story points, hours spent in the seat, hours spent per task, bugs fixed per week, and who knows what else. None of these metrics seems to work, and they all end up getting gamed.
That’s why we aren’t doing ourselves any favors when we start doing things like “tokenmaxxing”.
First, of course, measuring “tokens burned” doesn’t really tell you anything. Second, if you actually make tokens burned a target, well, we know what happens. People will severely game the system, and then Goodhart’s law kicks in. It turns out that Meta developers — like the cobra farmers in India — were using tools like OpenClaw to burn through massive chunks of tokens for no purpose at all.
Third, and worst of all, the managers in the corner office might notice. The execs are on an eternal quest for the best way to measure developers, and if they see this dashboard, they might actually latch on to the idea. And once that happens, things will go south very quickly.
The new lines of code
Token use is easy to count, looks great on a dashboard, and is utterly useless for anything at all except seeing how much electricity was jolted through GPUs. I can see the OKRs forming in the minds of executives as I type this. The one thing we don’t want to see is a slide at an investor briefing breathlessly announcing “Token throughput is up 30% YoY!”
And just like with lines of code, it’s generally true that maximizing token consumption is actually a negative indicator for quality and success. Getting Claude Code to use up tokens can actually produce worse outcomes than closely managing resources and keeping the coding agent on task. We don’t want token usage to become the new busy work, with developers burning up resources and execs patting them on the back as usage charts climb up and to the right — all to no avail.
It wasn’t long before word got out, and Meta shut it down.
And they were smart to act quickly. There is no upside to tokenmaxxing — it’s a perverse incentive.
Tokenmaxxing is just lines of code dressed up in a tuxedo. Or better, a clown suit.
