According to a column by the New York Times’ Kevin Roose, employees at companies including Meta and OpenAI compete on “internal leaderboards that show how many tokens[…]each worker consumes.” At Meta in particular (and also Shopify), Roose says volume of A.I. used has become a metric that goes into people’s evaluations, with managers “rewarding workers who make heavy use of A.I. tools and chastening those who don’t.”
Analogies are tricky here. One is tempted to say it’s like making painters compete to use the most paint, but even if the paint is just being splattered as quickly as possible, it’s at least going to be visible when the project is done. It’s a bit more like telling soldiers to gauge their battlefield success by the number of bullets fired, but suppressive fire that doesn’t hit anything has its place in war strategy. The best analogy I can come up with is this: it’s like NBA mascots being evaluated by how many t-shirts they fire out of their t-shirt cannons, but the t-shirts are made by Hermès.
The resulting numbers, in terms of both tokens and money, are absolutely staggering. One OpenAI engineer, according to Roose, burned through 210 billion tokens, which Roose equates to 33 Wikipedias. A Swedish software engineer claims to Roose that his company spends more than his salary on his Claude Code tokens alone.
This “tokenmaxxing” trend clearly stems in part from the use of “claws,” agentic AI platforms like OpenClaw, which are this year’s biggest supposed innovation in AI. OpenClaw’s virality was part of the big shift away from OpenAI’s GPT models and toward Claude this year by AI fanatics, and OpenAI subsequently hired OpenClaw’s creator, seemingly in a bid to maintain its position as the industry leader.
But even when used without an external claw platform Claude Code is becoming more and more like OpenClaw lately, with a feature rolling out last week that allows greater and greater on-the-go vibe coding, by letting users communicate with Claude Code more easily on their phones.
We just released Claude Code channels, which allows you to control your Claude Code session through select MCPs, starting with Telegram and Discord.
Use this to message Claude Code directly from your phone. pic.twitter.com/sl3BP2BEzS
— Thariq (@trq212) March 19, 2026
The promo even features a little 4-bit sprite of a lobster, or possibly a crab—a red crustacean at any rate—the new symbol for LLM token profligacy.
But tokenmaxxing speaks to a wider issue in which these companies tout the sheer number of tokens processed as a marker of success. OpenAI president Greg Brockman bragged about a week ago that the coding-oriented GPT-5.4 processes 5 trillion tokens per day—which, in fairness, makes sense as a way to please investors, because tokens cost money.
gpt-5.4 has ramped faster than any other model we’ve launched in the API: within a week of launch, 5T tokens per day, handling more volume than our entire API one year ago, and reaching an annualized run rate of $1B in net-new revenue.
it’s a good model, try it out!
— Greg Brockman (@gdb) March 16, 2026
But 5 trillion is a huge number, you have to admit. Did you know that Ronald McDonald’s Big Red Shoe car in the Macy’s Thanksgiving Day Parade would fit a men’s size 266 foot if it were to have an actual foot in it? Aren’t big numbers cool?
Read the full article here
