Tokenmaxxing is over at Meta, at least on paper. On 3 September 2026 the company confirmed that performance evaluations will no longer be tied to how much artificial intelligence an employee consumes, ending roughly ten months in which staff were graded on “AI-driven impact” and labelled AI Native, AI First or AI Enabled according to how heavily they leaned on chatbots and agents. A Meta spokesperson, Tracy Clayton, put it plainly: the company “will not use AI adoption dashboards or token counts to evaluate impact.”
In the same week, Meta began pushing an agentic tool called Hatch onto employee machines. Hatch browses the web, drives applications and completes multi-step tasks without supervision, and an agent of that shape consumes far more tokens than the chatbot it replaces. So the company has removed the scoreboard while handing everyone a faster engine. That contradiction is the story, and it is a preview of the decision every business running autonomous AI agents will face within the next two quarters.
This article sets out what actually changed, what it cost, and what the retreat from tokenmaxxing means for anyone budgeting AI spend outside Menlo Park. Our AI models and tools hub tracks these shifts as they land. Every figure below is attributed to the outlet that reported it, and where sources disagree, we say so rather than picking the number that reads best.
Table of contents
- What Meta Actually Changed About Tokenmaxxing
- Hatch, the Agent Meta Is Pushing as Tokenmaxxing Ends
- Claudeonomics: How Tokenmaxxing Got a Leaderboard
- What Tokenmaxxing Actually Cost
- Tokenmaxxing Was Never a Meta-Only Problem
- Why Tokenmaxxing Produced Motion Instead of Progress
- The Contradiction at the Centre of the Tokenmaxxing Retreat
- What Replaces Tokenmaxxing at Meta
- What the Tokenmaxxing Reversal Means for Everyone Else
- What We Could Not Verify About the Tokenmaxxing Story
- Tokenmaxxing at Meta: Common Questions
- References and Further Reading
What Meta Actually Changed About Tokenmaxxing
The change is narrow, specific and easy to overstate. Meta did not ban AI, cap accounts or switch anything off. It edited the wording of a performance review framework, and that edit removes the mechanism that made tokenmaxxing rational in the first place.
The sentence that did the work
Meta introduced AI-driven impact as a core expectation in November 2025. Reviewers were asked to assess how far an employee’s results were produced with AI assistance, and internal tooling made token consumption visible alongside the answer. The revised guidance replaces that test with a much looser one: outcomes “can be supported by AI or other means.” Nine words end the tokenmaxxing incentive, because the moment a manager cannot see your token count, inflating it earns nothing.
What Meta says it will no longer measure
Three things are named explicitly in the reporting: AI adoption dashboards, raw token counts, and the AI Native, AI First and AI Enabled labels attached to individuals. The Information reported the change on Wednesday, and Meta confirmed it. What has not been withdrawn is the expectation that engineers ship faster, or Mark Zuckerberg’s standing request that Meta’s codebases be rewritten so AI agents can read and modify them more easily.
The tokenmaxxing timeline
The sequence matters more than any single announcement, because it shows a company reversing an incentive it built deliberately and defended publicly.
| Date | Event |
|---|---|
| November 2025 | AI-driven impact becomes a core review expectation |
| March 2026 | Kevin Roose’s New York Times column names the behaviour publicly |
| 9 April 2026 | The Information reveals the Claudeonomics leaderboard |
| 11 April 2026 | Leaderboard shut down, two days after the report |
| April 2026 | Keystroke collection for AI training announced |
| May 2026 | Development of the Hatch agent begins |
| June 2026 | Keystroke project paused; Google throttles Meta’s token supply |
| 12 June 2026 | Bosworth memo reaches roughly 6,000 staff |
| 2 September 2026 | Meta moves from Google Chat to Slack to host AI agents |
| 3 September 2026 | Token counts formally dropped from performance reviews |
Hatch, the Agent Meta Is Pushing as Tokenmaxxing Ends
Hatch is the reason this week’s announcement is not a simple retreat. Meta is withdrawing the measurement while accelerating the behaviour the measurement was supposed to encourage, and it is doing so with a tool built for exactly the kind of long-running, high-consumption work that made tokenmaxxing expensive.
What Hatch actually does
Hatch is an agentic platform in the mould of OpenClaw. It takes an instruction, opens a browser, operates applications and works through a task without a human in the loop. Development started in May 2026 and employees have been running it on corporate devices for several weeks ahead of a wider release. Some have used it for personal errands, including booking appointments, which tells you how general the tool is meant to be.
Agents of this class are trained with reinforcement learning on long-horizon tasks, which is precisely why they keep working long after a chat session would have stopped. That design choice is the root of the cost problem tokenmaxxing merely made visible.
Why an agent changes the token arithmetic
A chatbot session is bounded by a person’s typing speed. An agent is bounded by its task. Meta’s AI chief, Alexandr Wang, has said that “everyone at the company will benefit from a robust and useful agent ecosystem,” and he is probably right about the capability. He is also describing a workload that consumes tokens continuously rather than in bursts. Removing the leaderboard does not reduce consumption when the replacement tool runs unattended for hours.
Why some employees are hesitant
Adoption is voluntary, and not everyone is volunteering. The stated hesitation is privacy: Hatch works best when connected to accounts and given broad device access, and Meta announced in April 2026 that it intended to collect employee keystrokes as AI training data. That project was paused in June, but the memory of it is doing real work in staff rooms. Meta’s move from Google Chat to Slack on 2 September, made specifically to host agents, has sharpened the same concern.
Claudeonomics: How Tokenmaxxing Got a Leaderboard
Nobody at Meta was ordered to compete on token consumption. An employee built a dashboard, the dashboard became a game, and the game became a performance signal. That progression is the clearest illustration of how tokenmaxxing takes hold.
A dashboard for 85,000 people
The tool was called Claudeonomics, named after Anthropic’s Claude. It ranked more than 85,000 employees by monthly token consumption and displayed the top 250 publicly, awarding titles such as Token Legend and Cache Wizard. Neither Zuckerberg nor chief technology officer Andrew Bosworth appeared in the top 250. The leaderboard was informal, unofficial and, for a few months, one of the most consequential pieces of software inside the company.
The consumption figures
Two totals have been reported for the same broad window, and the gap between them is itself informative. The April reporting put employee consumption above 60.2 trillion tokens in 30 days. A later figure, cited as the reading before the leaderboard came down, was 73.7 trillion. As a share of the higher number, the earlier total sits at about 82 per cent, which is the growth rate that alarmed finance long before tokenmaxxing became a headline.
The two days it lasted after exposure
The Information published on 9 April 2026. Two days later the dashboard was gone, replaced by a notice from its creator: “It was meant to be a fun way for people to look at tokens, but due to data from this dashboard being shared externally, we’ve made the decision to shutter Claudeonomics for now.” Meta says it did not ask for the removal. The tokenmaxxing behaviour outlived the dashboard by roughly five months.
What Tokenmaxxing Actually Cost
Estimating the bill is harder than it looks, because nobody outside Meta knows the blend of models, the cache hit rate or the negotiated rate. Published estimates for the same month differ by a factor of four, and it is worth understanding why before quoting any of them.
The two public estimates
One widely cited figure puts the 60 trillion tokens at roughly $221 million for the month. Another analysis, applying standard published API rates, arrives at about $900 million. Both are defensible arithmetic on different assumptions. Enterprise contracts, cached input pricing and cheaper models pull the number down; frontier reasoning models pull it up. If you are modelling your own exposure, our AI token cost calculator walks through the same assumptions at a scale you can actually verify.
One person, $1.4 million
The single most quoted number from the leaderboard is not the company total. The top-ranked individual averaged 281 billion tokens, which at the cheapest available Claude Opus 4.6 rate of $5 per million tokens works out at about $1.4 million for one employee in one month. That is a headcount-scale cost attached to a single seat, and it is the figure that turned tokenmaxxing from a curiosity into a board-level question.
Against a $145 billion capital plan
Context cuts both ways. Meta planned roughly $145 billion of spending in 2026, most of it AI infrastructure. Against that, even the $900 million estimate is a rounding error. But internal token spend is operating expenditure on someone else’s model, not capital that builds an asset, and Meta was also reported by the Financial Times in June 2026 to have had its token supply throttled by Google. Scarcity, not cost alone, is what forces rationing.
Tokenmaxxing Was Never a Meta-Only Problem
The research firm SemiAnalysis coined the term to describe organisations that encouraged heavy AI usage before putting financial controls in place. By that definition, most of the industry qualified during the first half of 2026.
The comparison across five companies
Each company reached the same wall by a slightly different route, and the differences are instructive when you are deciding which control to copy.
| Company | Mechanism | Outcome |
|---|---|---|
| Meta | Employee-built leaderboard plus review criteria | Dashboard removed April, criteria dropped September |
| Microsoft | Internal token leaderboard since January 2026 | Senior staff top the ranks; engineers admit inflating usage |
| Salesforce | Minimum monthly spend floors, caps removed | Peer comparison made visible to reduce friction |
| Shopify | First leaderboard, 2025, later rebranded | Renamed a usage dashboard; circuit breakers added |
| Amazon | Usage expectations in performance management | Agents spun up on meaningless tasks to hold stats |
| Uber | Annual token budget, no per-seat throttle | Full-year 2026 budget exhausted in four months |
Salesforce set a floor rather than a ceiling
Salesforce is the clearest inversion of the usual control. Rather than capping spend, it set minimum monthly targets of $100 for Claude Code and $70 for Cursor, made peer comparison visible, and removed maximum caps entirely to take friction out of development. Expressed against the Claude Code floor, the Cursor floor sits at 70 per cent. A mandated minimum is tokenmaxxing as policy rather than as folklore.
Uber ran out of road first
Uber burned its entire 2026 token budget in the first four months of the year, driven substantially by Claude Code. Chief operating officer Andrew Macdonald has said the company struggled to connect measurable productivity gains for individual workers with any company-wide impact. That gap between local speed and organisational output is the same one described in our analysis of why workforce planning models are not keeping up with AI.
Why Tokenmaxxing Produced Motion Instead of Progress
The strongest argument against tokenmaxxing did not come from a critic. It came from Meta’s own chief technology officer, in a memo sent to roughly 6,000 employees in June 2026.
Bosworth’s memo
Andrew Bosworth wrote: “Nobody should be using AI tools just for the sake of using them. All motion is not progress and token usage alone is not a measure of impact of any kind.” It is an unusually direct repudiation of a metric his own company had normalised, and it landed three months before the review criteria actually changed. The delay is the interesting part; culture moved slower than the memo.
What engineers admitted doing
The confessions are specific. A Microsoft engineer described asking AI questions already answered by existing documentation, prototyping features with no intention of shipping them, and defaulting to an agent even when the manual approach was faster, all to avoid appearing insufficiently AI-native. At Amazon, employees reportedly spun up agents on tasks that served no purpose beyond holding their statistics up. Tokenmaxxing is not laziness; it is a rational response to a badly chosen measurement.
The code quality cost
Meta developers described massive waste, with agents burning tokens for minimal output and careless generated code causing site events. This is the part rarely captured in a token bill, because the cost lands later as incident response, review time and rework. An organisation that rewards consumption gets consumption, and the cleanup appears in a different budget line from the one being optimised.
The Contradiction at the Centre of the Tokenmaxxing Retreat
Read the two announcements together and they do not obviously agree. Meta has stopped measuring AI usage and started distributing a tool designed to increase it. Both can be sincere, and the reconciliation matters for anyone writing an internal AI policy this quarter.
Removing the metric is not removing the pressure
Formal criteria are only part of how expectations travel. Zuckerberg’s request that codebases be rewritten for agent comprehension still stands, and Reuters reported that Meta had planned workforce reductions of up to 60 per cent in certain areas, a plan that stalled partly because AI-generated software was too buggy. An employee reading those signals will not conclude that AI enthusiasm has stopped counting, whatever the review template now says.
The lawsuit in the background
There is a legal dimension that rarely makes the headline. Around two dozen former employees have brought a claim alleging that the AI-usage metrics unfairly penalised staff who were on health or family leave, and therefore consuming fewer tokens through no fault of their own. Any metric that scales with hours at a keyboard will disadvantage people who are not at one. That is a governance flaw, not an implementation detail, and it is a strong reason to drop a usage metric regardless of cost.
Voluntary adoption is still adoption
Meta’s position is that Hatch usage is optional. In practice, an agent that visibly makes colleagues faster generates its own pressure, and the token consumption arrives whether the leaderboard exists or not. The difference is that the spend is now attributable to work rather than to scoreboard climbing, which is exactly the distinction the new controls are designed to enforce.
What Replaces Tokenmaxxing at Meta
Dropping a metric leaves a vacuum, and Meta is filling it with plumbing rather than exhortation. Three mechanisms have been reported, and together they describe the standard enterprise pattern for the next eighteen months.
The AI Gateway and structured budgets
Meta has built a central dashboard, the AI Gateway, to track internal usage and spend in one place, with automatic alerts for unusual cost spikes. Structured token budgets are expected to arrive in 2027. Note what this is: not a ban, but observability first and allocation second. Tokenmaxxing thrived precisely because consumption grew exponentially with, in the reporting’s phrase, little visibility or control.
MetaCode over third-party models
The second lever is substitution. Meta intends to steer engineers towards MetaCode, its in-house coding assistant formerly known as Devmate, and reduce reliance on third-party tools including Anthropic’s Claude, while keeping other models available. Moving inference in-house converts a per-token operating cost into infrastructure the company already owns, which is a different economic problem from the one tokenmaxxing created. Routing economics of this kind are why an AI token exchange has become a category worth watching.
Cost per accepted task
The third lever is the measurement that should have been there from the start. The metric that survives scrutiny is cost per accepted unit of work: merged pull requests, resolved tickets or shipped features per dollar of AI spend. It is harder to compute than a token count and impossible to game by leaving an agent running overnight, which is precisely the point.
| Metric | What it rewards | How it is gamed |
|---|---|---|
| Tokens consumed | Volume of model calls | Idle agents, redundant prompts, throwaway work |
| Seats licensed | Procurement coverage | Unused licences counted as adoption |
| Sessions per week | Habit formation | Trivial sessions opened to hold a streak |
| Merged pull requests per dollar | Reviewed, accepted output | Splitting work into smaller commits |
| Tickets resolved per dollar | Closed customer problems | Premature closure, reopened tickets |
| Cost per shipped feature | Delivered business value | Hard to game, slow to measure |
What the Tokenmaxxing Reversal Means for Everyone Else
Very few organisations have 85,000 employees or a $145 billion capital plan. The mechanism that produced tokenmaxxing, however, needs neither. It needs a visible usage number and a performance conversation, and most companies now have both.
The budgets that have emerged
Reported practice has converged quickly. Many organisations now run monthly per-employee AI budgets in the $250 to $500 range, while firms including Workday and Stripe allow around $2,000 per person per month. Against that $2,000 ceiling, a $500 budget is a quarter and a $250 budget an eighth. Limits are increasingly tiered by seniority and role, and AI is treated as an operating expense alongside software subscriptions and travel.
Four controls worth copying
Shopify’s circuit breakers are the most transferable idea here: automatic interruption of a runaway agent before it burns a month of budget on an infrastructure bug. Add central observability before allocation, so you know your baseline. Tier budgets by role rather than applying one number. And never publish an individual ranking, because a ranking is what converts a diagnostic into tokenmaxxing.
What to change in your next review cycle
If your performance framework mentions AI adoption, usage or tool coverage, replace it with an outcome the business already recognises. Meta’s revised wording, that outcomes can be supported by AI or other means, is a serviceable template. The organisations getting real returns are pairing that with genuine enablement, which is closer to the forward-deployed engineering model than to a dashboard, and with a deliberate AI strategy that names the work being changed.
What We Could Not Verify About the Tokenmaxxing Story
Reporting on internal tooling depends on people describing systems they are not supposed to describe. Several figures in this story deserve a caveat, and pretending otherwise would be dishonest.
Numbers that vary by source
The token totals differ: 60.2 trillion appears in April reporting, 73.7 trillion in later accounts, and it is not always clear whether the windows align. The cost estimates diverge further, from about $221 million to roughly $900 million for a comparable month. We have quoted both rather than averaging them, because the difference reflects genuinely unknown pricing rather than measurement error.
What Meta has not said
Meta has not published token totals, confirmed a cost figure, released the Hatch roadmap or given a date for the 2027 budgets. It has not said whether MetaCode will become mandatory, and it has not commented in detail on the former employees’ legal claim. The absence of tokenmaxxing from the review template is confirmed; almost everything about what replaces it is reported rather than announced.
What to watch next
Three signals will tell you whether this is a real reversal. Whether the AI Gateway budgets arrive on schedule in 2027. Whether Hatch ships externally, which would change Meta’s incentives entirely. And whether any large employer publishes a cost-per-accepted-task figure, which would move the industry past tokenmaxxing and its mirror image, token minimising, to something that actually measures value. As with all agent deployments, the governance question described in our piece on execution governance will matter more than the spend.
Tokenmaxxing at Meta: Common Questions
What does tokenmaxxing mean?
Tokenmaxxing is the deliberate inflation of AI usage to signal productivity or enthusiasm, typically by consuming as many model tokens as possible. The research firm SemiAnalysis coined the term for organisations that encouraged heavy usage before introducing financial controls.
Has Meta banned AI usage?
No. Meta has removed token counts and adoption dashboards from performance evaluation and retired the AI Native, AI First and AI Enabled labels. Employees retain access to AI tools, and the company is actively promoting its new Hatch agent.
What is Hatch?
Hatch is Meta’s agentic AI platform, developed from May 2026 and comparable in shape to OpenClaw. It performs computer tasks autonomously, including browsing the web and operating applications, and has been in employee testing on corporate devices for several weeks.
How much did tokenmaxxing cost Meta?
No official figure exists. Reported estimates for a single 30-day period range from about $221 million to roughly $900 million, against consumption of 60.2 to 73.7 trillion tokens. The top individual on the leaderboard averaged 281 billion tokens, worth about $1.4 million at the cheapest cited rate.
What was Claudeonomics?
Claudeonomics was an employee-built internal leaderboard, named after Anthropic’s Claude, that ranked more than 85,000 Meta staff by monthly token use and displayed the top 250 with titles such as Token Legend and Cache Wizard. It was shut down two days after The Information reported on it.
What should replace token counts as a metric?
Cost per accepted unit of work: merged pull requests, resolved tickets or shipped features per dollar of AI spend. It is slower to compute than a token total but cannot be inflated by leaving an agent running, which is what made tokenmaxxing possible.
Is tokenmaxxing happening at other companies?
Yes. Microsoft has run an internal token leaderboard since January 2026, Salesforce sets minimum monthly spend floors, Shopify built the first leaderboard in 2025, and both Amazon and Uber have publicly hit the limits of unbounded usage.
References and Further Reading
WIRED on Meta’s Agent Push and Tokenmaxxing
Fortune on the Claudeonomics Shutdown
The Decoder on Meta’s Shift to Token Managing
The State of AI on Meta Capping Employee Usage
Gizmodo on Meta’s Mixed Messages
The Pragmatic Engineer on the Trend
Fortune on the Missing Return on Investment
Fortune on Uber’s Exhausted Token Budget
The New York Times on Token Minimising
IBTimes on Enterprise AI Budgets
Techloy on AI Usage as a Performance Metric
IBM on Measuring Value Instead of Volume
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.