The Pulse: 'Tokenmaxxing' — How Big Tech Devs Are Burning AI Tokens to Hit Metrics

Source: The Pragmatic Engineer · Author: Gergely Orosz · Date: Apr 23, 2026 · Original article

---

What is "tokenmaxxing"?

A "token" is the unit of text an AI model processes — roughly a chunk of a word. When you prompt Claude, ChatGPT, or Cursor, you pay (or your employer pays) per token sent and received. More tokens = more money spent on the AI vendor's API.

Tokenmaxxing is a new trend at large tech companies where engineers deliberately burn as many tokens as possible — not to do better work, but to look like a heavy AI user on internal dashboards and leaderboards. It's Silicon Valley's newest form of conspicuous consumption: instead of flashy cars, it's flashy AI bills.

The pattern is the same one seen many times in software: a manager picks a number that correlates with productivity (lines of code, commits, now tokens), turns it into a target, and developers immediately learn to game it. This is Goodhart's Law in action — "When a measure becomes a target, it ceases to be a good measure."

This issue dives into how that's playing out at four big companies: Meta, Microsoft, Salesforce, and Shopify (the only one that seems to have avoided the trap).

---

Meta: the "Token Legend" leaderboard

An engineer inside Meta built an internal leaderboard — nicknamed "Claudeonomics" — that ranks all 85,000+ Meta employees by how many AI tokens they burn. Top users earn titles like Session Immortal and Token Legend. The Information broke the story.

What engineers inside Meta told Gergely about the reality on the ground:

• Massive waste. Many devs run an internal agent (referred to as "OpenClaw-like") that chews through huge token volumes for little to no useful output.
• Outages from AI overuse. Some SEVs (Meta's term for production incidents) appear to have been caused by careless AI-generated code being merged. The dev behind the SEV seemed more focused on volume of code than on quality.
• The leaderboard top is throwaway work. Meta has a tool called "Trajectories" that lets anyone view another employee's actual AI prompts. Looking at the prompts of the top-ranked tokenmaxxers makes it obvious the work is junk.

The numbers are staggering: Meta employees burned 60.2 trillion tokens in 30 days. At Anthropic's public API prices, that would be ~$900M/month. Even with bulk-discount enterprise pricing, the bill is likely $100M+/month — much of it from pure tokenmaxxing.

After The Information's story and the social-media backlash, Meta took the leaderboard down.

Was the leaderboard actually a training-data harvester?

A long-tenured Meta engineer floated a theory worth taking seriously: the leaderboard was never really about productivity. It was about generating AI usage traces — recorded sequences of prompts, tool calls, and code edits — that Meta can use as training data for its next-generation coding model. More usage = more traces = better training corpus.

In other words, Meta may have been willing to pay nine figures a month to manufacture training data, while letting employees believe they were chasing prestige. As the engineer put it: "if any company has the means to do so, it's Meta."

---

Microsoft: full-force tokenmaxxing

Microsoft has run an internal token leaderboard since January. Two oddities surfaced almost immediately:

• Distinguished engineers — the most senior individual contributors, who historically write very little code — sit in the company-wide top 5.
• VPs, who spend most of their day in meetings, made the top 10–20.

That alone tells you the metric isn't measuring "engineering productivity." It's measuring "willingness to pipe everything through an LLM."

A relatively new Microsoft engineer told Gergely they tokenmaxx out of fear, not vanity. Performance reviews and dashboards now track AI usage, token usage, and percentage of code written by AI vs. by hand. They don't want to be tagged as "uses too little AI" — so they deliberately inflate their numbers. Their tactics:

• Ask the AI questions whose answers are already in the docs. The AI fetches and chews through the docs to give an answer 10× slower than just opening the internal "readthedocs" tool — but readthedocs doesn't count toward token metrics, while AI does.
• Prompt the AI to prototype features they have no intention of shipping. A few back-and-forths, then throw it all away.
• Default to the agent even when hand-coding is faster. Hit "go," watch it fail, repeat.

This is a new hire actively making themselves slower to avoid being seen as not-AI-native. That's the kind of behavior the metric is creating.

---

Salesforce: a "minimum spend" target

Salesforce has gone further — it has effectively turned token spend into a floor, not just a leaderboard. Three internal tools shape the incentives:

1. A Mac menu-bar widget that updates every 15 minutes showing your personal token spend, alongside the company's "minimum expected spend." Recent targets: $100/month on Claude Code, $70/month on Cursor.
2. A web tool to look up any colleague's token spend — used to peek at what teammates are doing.
3. A monthly "maximum" ($250 Claude Code, $170 Cursor) that, until recently, capped spend. The cap can be exceeded with one click. Last week, some Salesforce engineering orgs had the maximum removed entirely to "remove friction from the development process."

The signal to staff is unambiguous: spend at least $170/month on tokens or get flagged. Predictably, devs game it:

• Pure busywork. Asking Claude or Cursor to "build me X" where X is a side project unrelated to their actual work — purely to burn tokens.
• Calibrating to slightly above average. Devs check peers' spend, find the just-above-average sweet spot, and aim for it. Not too low (suspicious), not too high (wasteful and visible).

---

Shopify: how to do this without the gaming

Shopify built what's likely the first-ever token leaderboard, back in 2025 — and it actually worked. The Head of Engineering, Farhan Thawar, explained the original intent on The Pragmatic Engineer Podcast last June:

"We have a leaderboard where we actively celebrate the people who use the most tokens because we want to make sure they're celebrated if they're doing great work with AI. [For top users,] I want to see why they spent $1,000 a month in credits for Cursor. Maybe it's because they're building something great with an agent workforce underneath them!"

A year later, Farhan shared the updates with Gergely. Three things make Shopify's setup different:

1. They renamed the "leaderboard" to a "usage dashboard." This is a small wording change with a big psychological effect — it stops framing token use as a competition. Token spend is shown on internal wiki profiles too, so it's transparent without being a podium.
2. Circuit breakers for runaway agents. If your personal spend spikes within a single day, access is automatically cut off. You can re-enable it if the spike was intentional. This catches a real problem: an agent that gets stuck in a loop and quietly burns thousands of dollars overnight. Bonus side effect — the circuit breaker has surfaced actual infrastructure bugs.
3. High usage triggers a conversation, not a trophy. Farhan personally checks in with the top-spending engineers to understand what they're working on. That single human-in-the-loop step makes tokenmaxxing pointless and embarrassing — if you're at the top because you were running garbage prompts, that conversation is going to be awkward.

Farhan's most interesting observation: the interesting metric isn't "who spent the most overall?" — it's "whose individual tokens cost the most?" Tokens get expensive when you feed in long, complex contexts (large codebases, deep documents, careful reasoning chains). Engineers generating expensive individual tokens turn out to be doing genuinely deep work, not high-volume slop.

---

Why tokenmaxxing is bad for everyone except AI vendors

Gergely's bottom line: there's almost no rational reason for a company to incentivize raw token usage. The downsides compound:

• Direct cost. Token spend balloons without matching value — sometimes tens of millions of dollars a month.
• Slower work. As the Microsoft example shows, devs deliberately route around faster internal tools to keep token numbers up.
• Busywork as a strategy. Engineers prompt projects they have no intention of shipping, just to burn tokens.
• Quality risk. As at Meta, "ship lots of AI code" mindsets have been linked to production incidents.

The cleanest analogy: tokens are the new lines-of-code metric. There was a time when companies measured productivity by lines of code per day or per month. It collapsed quickly once managers realized two things:

1. Lines of code are trivially gameable — write boilerplate, copy-paste, split files.
2. The best engineers often write less code: they solve hard business problems quickly and reliably, sometimes with no code at all.

Tokens are exactly the same trap, with one twist: gaming lines of code costs nothing, but gaming token counts produces a massive AI bill paid directly to Anthropic, OpenAI, and the rest. Tokenmaxxing is great for AI vendors. For the company paying the bill, it's expensive theater.

---

The bigger picture

This piece is a free excerpt of a larger Pulse issue. The full issue also covers:

• Token spend breaking budgets at 15 different companies, and how engineering leaders are scrambling to respond.
• AI vendors capping individual users. GitHub Copilot and Anthropic are starting to throttle lower-revenue individual subscribers to free up capacity for business customers, whose spend has 10×'d in months. OpenAI/Codex is the exception so far.
• Morale at Meta reportedly hitting an all-time low, with looming layoffs and an invasive tracking program rolled out to all US employees.

This summary covers the free, public excerpt of the issue. The full Pulse with the additional sections is paywalled.