From Tokenmaxxing to Tokenminimizing: Meta Moves to Cap Employee AI Usage

Meta is reining in the very behavior it spent the past year encouraging. The company plans to impose limits on employees' AI token usage to control costs that are projected to run into the billions in 2026, according to an internal memo first reported by The Information's Jyoti Mann on Thursday.

The memo, which The Information reviewed and which Meta shared with roughly 6,000 staffers earlier this week, describes a broader efficiency program. According to the report, Meta told employees it has seen exponential growth in internal AI usage while individuals and teams have had little visibility into or control over their own consumption, and that in 2027 the company expects to manage tokens in a more structured way, with budgets, allocation decisions, and supporting tools.

To get there, a team of product developers and engineers has built a central dashboard called AI Gateway that monitors usage and spending in real time, per The Information. The company plans to roll out automated alerts for unusual spending spikes, and the group is tracking current costs to forecast future spend, plan computing capacity, and negotiate with vendors. Meta intends to announce the new controls to a wider set of employees in the coming weeks.

Notably, the report says Meta will also nudge staff away from third-party AI tools and toward in-house alternatives, including its internally developed coding assistant MetaCode, previously known as Devmate, while still allowing access to new external models. Engineers in Meta's recently formed Applied AI Engineering division have reportedly been tasked with improving MetaCode to reduce the company's heavy reliance on Anthropic's Claude for coding work, including by generating programming challenges to create reinforcement learning training data.

A Meta spokesperson told The Information that controlling AI costs is a known priority and that the company remains focused on using AI to help employees in their daily work.

A whiplash reversal

The pivot is striking given where Meta was just two months ago. The company spent months aggressively pushing AI adoption, giving staff access to its own models alongside tools from OpenAI, Anthropic, and Google, and telling employees last November that demonstrating AI-driven impact would be a core expectation this year.

That push produced some unintended consequences. In April, The Information revealed that employees were competing on an internal leaderboard called Claudeonomics, which ranked the company's top 250 token users and fueled a status game known as tokenmaxxing. Some staffers inflated their numbers by instructing AI agents to run multiple tasks at once. A copy of the dashboard reviewed by The Information showed 60.2 trillion tokens consumed in a 30-day period, a figure that later climbed to 73.7 trillion before the leaderboard came down.

Even before this week's memo, CTO Andrew Bosworth had been trying to cool things off. In an April memo reported by The Information, he told staff that "nobody should be using AI tools just for the sake of using them," arguing that motion is not progress and that token usage alone measures nothing about impact. The same memo unveiled an initiative called the Agent Transformation Accelerator, aimed at unifying what Bosworth described as a fragmented internal AI tooling landscape.

Why it matters

Meta is not alone here. The Information has also reported that Uber and ServiceNow exhausted their entire annual budgets for Anthropic's tools within the first few months of 2026, and that some venture firms are capping employee usage after individual staffers racked up thousands of dollars a day in token costs.

The tokenmaxxing-to-tokenminimizing arc captures the broader story of enterprise AI in 2026. Companies that treated raw usage as a measure of transformation are discovering that tokens are an input metric, not an output metric, and an expensive one at that. When usage becomes the scoreboard, employees optimize for the scoreboard.

The tension is sharpest at Meta, which plans to spend as much as $145 billion this year, much of it on AI infrastructure, even as it clamps down on what its own employees spend using AI. With investors pressing for returns, the company is leaning on new revenue streams like paid subscription tiers across Facebook, Instagram, and WhatsApp and planned charges for its AI business agents. The open question is whether limiting usage forces a more honest accounting of which AI work pays for itself, or simply undercuts the productivity gains the whole spending spree was meant to deliver.