Meta drops AI usage from performance reviews after "tokenmaxxing" backfires
Meta is removing AI tool usage from engineer performance reviews. AI dashboards and token counters will no longer factor into how engineers are rated, according to an internal memo from executives Maher Saba and Santosh Janardhan, seen by The Information. The memo states that quality, speed and complexity of the work are what count now.
The change reverses an earlier policy that made AI usage a review criterion, which employees turned into what became known internally as "tokenmaxxing": burning through large volumes of AI tokens simply to look good on internal leaderboards, without that effort translating into better output. The behavior carried a real cost. Meta's internal AI use alone is heading toward billions of dollars in spending in 2026. In response, Meta plans to introduce usage budgets and a central dashboard for AI spending starting in 2027.
Separately, Meta is testing a new AI agent tool called Hatch, which is designed to handle computer tasks on its own, according to a WIRED report. The tests have run into internal pushback: some employees are reluctant to connect Hatch to their personal accounts because of privacy concerns.
Key facts
- Meta will no longer count AI dashboard metrics or token counters toward engineer performance reviews.
- Executives Maher Saba and Santosh Janardhan said in an internal memo that quality, speed and complexity of work now matter instead.
- The old AI-usage metric led to "tokenmaxxing": employees running up token counts to rank well on internal leaderboards.
- Meta's internal AI use is heading toward billions of dollars in costs in 2026, prompting plans for usage budgets and a central dashboard starting in 2027.
- Meta is separately testing an AI agent tool called Hatch for handling computer tasks, and some employees are wary of connecting it to personal accounts over privacy concerns.
Why it matters
It is a public admission that a widely used proxy for AI adoption, how much staff use AI tools, measured the wrong thing. Rewarding token volume did not reward better engineering; it rewarded gaming the leaderboard, and it ran up a real bill. The reversal is a data point for any organization tempted to turn AI usage itself into a KPI.
Who it affects
Meta's engineering staff, whose reviews now drop the AI-usage criterion in favor of output quality, speed and complexity. It also affects Meta's finance and infrastructure planning, since the company is now moving to budget and centrally track internal AI spending starting in 2027. Employees asked to test the Hatch agent tool are affected as well, particularly around the personal-account access it requests.
How to use it
There is no product or pricing here for readers to act on. The practical takeaway is for other companies running AI-adoption metrics: tying reviews or incentives to raw AI usage invites gaming rather than genuine productivity gains, and usage volume is worth budgeting and tracking centrally rather than left as an open-ended perk.
How solid is it
The core claim, that Meta dropped AI usage as a review factor, rests on an internal memo from named executives, Maher Saba and Santosh Janardhan, that The Information reviewed directly. The Hatch agent-testing detail comes from a separate WIRED report. Both are secondary citations relayed through The Decoder rather than the memo or WIRED report themselves.
Risks and caveats
The source gives no specific dollar figure for the 2026 internal AI costs, only that they are heading toward billions, and no date for how long the tokenmaxxing practice ran before Meta acted. It also does not specify what "computer tasks" Hatch performs beyond that general description, nor how many employees are declining to connect it to personal accounts.
“AI dashboards and token counters won't factor into performance reviews anymore.”
— Maher Saba and Santosh Janardhan, in an internal Meta memo seen by The Information