--- title: "Meta Stops Grading Engineers on AI Token Counts and What" canonical: "https://www.metaintro.com/blog/meta-drops-ai-token-counts-performance-reviews-2026" language: "en" author: "drashtigarach" published: "2026-09-03T11:52:11.000Z" modified: "2026-10-02T19:46:46.599Z" --- [Back to Blog](/blog) [AI](/blog/tag/ai)[Technology](/blog/tag/technology)[Career](/blog/tag/career)[Jobs](/blog/tag/jobs)[News](/blog/tag/news) # Meta Stops Grading Engineers on AI Token Counts and What That Signals for Your Review Meta has removed AI usage from its performance review criteria after a year of pushing staff to maximise tokens. What the reversal signals for your review. [![Drashti Garach](https://cdn.metaintro.com/rs:fill:40:40/q:72/plain/images/5719d740-e510-42bc-8017-e040d145f35f_1766029465094.png)Drashti Garach @DrashtiGarach](/blog/author/drashtigarach) [September 3, 2026](/blog/archive/2026/09)12 min read ![Meta Stops Grading Engineers on AI Token Counts and What That Signals for Your Review](https://cdn.metaintro.com/rs:fill:1200:675/q:78/plain/images/10141r1-gpu-rack.png) [https://x.com/intent/tweet?text=Meta%20Stops%20Grading%20Engineers%20on%20AI%20Token%20Counts%20and%20What%20That%20Signals%20for%20Your%20Review&url=https%3A%2F%2Fwww.metaintro.com%2Fblog%2Fmeta-drops-ai-token-counts-performance-reviews-2026](https://x.com/intent/tweet?text=Meta%20Stops%20Grading%20Engineers%20on%20AI%20Token%20Counts%20and%20What%20That%20Signals%20for%20Your%20Review&url=https%3A%2F%2Fwww.metaintro.com%2Fblog%2Fmeta-drops-ai-token-counts-performance-reviews-2026)[http://www.facebook.com/sharer.php?u=https%3A%2F%2Fwww.metaintro.com%2Fblog%2Fmeta-drops-ai-token-counts-performance-reviews-2026](http://www.facebook.com/sharer.php?u=https%3A%2F%2Fwww.metaintro.com%2Fblog%2Fmeta-drops-ai-token-counts-performance-reviews-2026)[https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fwww.metaintro.com%2Fblog%2Fmeta-drops-ai-token-counts-performance-reviews-2026&title=Meta%20Stops%20Grading%20Engineers%20on%20AI%20Token%20Counts%20and%20What%20That%20Signals%20for%20Your%20Review](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fwww.metaintro.com%2Fblog%2Fmeta-drops-ai-token-counts-performance-reviews-2026&title=Meta%20Stops%20Grading%20Engineers%20on%20AI%20Token%20Counts%20and%20What%20That%20Signals%20for%20Your%20Review)[mailto:?subject=Meta%20Stops%20Grading%20Engineers%20on%20AI%20Token%20Counts%20and%20What%20That%20Signals%20for%20Your%20Review&body=https%3A%2F%2Fwww.metaintro.com%2Fblog%2Fmeta-drops-ai-token-counts-performance-reviews-2026](mailto:?subject=Meta%20Stops%20Grading%20Engineers%20on%20AI%20Token%20Counts%20and%20What%20That%20Signals%20for%20Your%20Review&body=https%3A%2F%2Fwww.metaintro.com%2Fblog%2Fmeta-drops-ai-token-counts-performance-reviews-2026) The company that did most to popularise measuring employees by how much AI they use has just stopped doing it. [WIRED](https://www.wired.com/story/meta-pushes-its-new-ai-agent-on-employees-but-eases-off-on-tokenmaxxing/) reported, citing Meta insiders, that newly published performance review guidance no longer makes usage of AI a criterion for assessment, and [The Information](https://www.theinformation.com/briefings/exclusive-meta-tells-engineers-ai-token-usage-part-performance-reviews) reported that engineers were told AI adoption dashboards and token counts would no longer form part of evaluations. If your own employer has been pushing you to demonstrate AI usage, this is the most useful data point of the year. At [Metaintro](https://www.metaintro.com), we read reversals like this as leading indicators rather than as company news. ## What exactly did Meta change? The change is to the language of the review itself, which is where these things actually bite. WIRED reported that the newly published guidance no longer lists usage of AI as an assessment criterion, and that when describing desired outcomes it now says those outcomes can be supported by AI or other means. That phrasing does a lot of work. AI remains encouraged, but it is no longer the thing being measured. The measurement moves back to the outcome, and the tool becomes one available route to it rather than the objective in itself. That is a meaningful distinction for anyone whose manager has been asking them to evidence AI adoption, because it changes what a good answer looks like from a description of activity to a description of a result. Alongside that, [Gizmodo](https://gizmodo.com/meta-is-reportedly-sending-mixed-messages-to-employees-on-tokenmaxxing-2000806644) reported that engineers had been told AI adoption dashboards and token counts would no longer be part of evaluations. Those two mechanisms are worth naming separately because they are what turned an encouragement into a pressure. A dashboard that displays individual usage, and a count that can be ranked, converts a suggestion into a target. Some of WIRED's sources inside the company said they were relieved not to feel under pressure to make constant use of AI, which tells you how the previous arrangement was actually experienced. Relief is a revealing word. It suggests the previous policy was producing effort directed at the metric rather than at the job, which is the cost of an input measure that few organisations account for when they introduce one. ## What was tokenmaxxing in the first place? The term entered general circulation through a March 2026 column by Kevin Roose in the [New York Times](https://www.nytimes.com/2026/03/20/technology/tokenmaxxing-ai-agents.html), which described AI enthusiastic workplaces including Meta attempting to maximise AI usage among staff. Roose reported that Meta was rewarding workers who made heavy use of AI tools and chastening those who did not, and described leaderboards that turned AI consumption into a competition between colleagues. Once usage is on a leaderboard it stops being a productivity measure and becomes a performance in the theatrical sense. Colleagues compare positions, people who are naturally cautious about a new tool look uncommitted, and the person who found a way to do the work in fewer steps appears to be contributing less than the person who took more. Efficiency and consumption point in opposite directions, and a leaderboard cannot tell them apart. That is the predictable failure mode of measuring an input rather than an output. If token count is what gets rewarded, the rational response is to generate tokens, whether or not the work needed them. Meta was not alone in this, and the broader pattern of [companies tracking employee AI usage and feeding it into performance reviews](https://www.metaintro.com/blog/companies-track-employees-ai-usage-performance-reviews) has been spreading across the industry, alongside cases where [AI usage targets turned out to be inflated](https://www.metaintro.com/blog/amazon-ai-usage-targets-inflated) once anyone examined them closely. The pattern repeats across industries and predates AI entirely. Any metric that is easy to influence and loosely connected to the outcome will be influenced rather than earned, and the organisations that learn this fastest are the ones that stop collecting it. What makes this instance notable is the speed of the reversal. The practice was named publicly in March and substantially unwound by September, which is unusually fast for a management fashion. ## Why is Meta easing off now? Two pressures appear to have converged. The first is cost and supply. In June, reporting indicated that Google was throttling Meta's AI tokens, forcing the company to reconsider how freely it could encourage consumption. That pressure produced a striking outcome, in which [Meta began limiting how much AI its own staff could use](https://www.metaintro.com/blog/meta-is-limiting-how-much-ai-its-own-staff-can-use-and-what-that-says-about-tech-jobs). An incentive scheme that rewards consumption becomes untenable when the supply of the thing being consumed is capped or expensive. There is a general lesson in that for employees. Mandates driven by abundance tend to soften when the economics change, so a policy introduced in a period of cheap capacity is rarely permanent. The second is that the measure was not producing what it promised. Input metrics rarely survive contact with a serious review of returns, because they cannot be tied to a business outcome. There is also a trust dimension. Meta faced a backlash in April when it emerged that the company planned to use employee keystrokes to train AI, a plan that turned [keystroke tracking into a public dispute with its own workforce](https://www.metaintro.com/blog/meta-employee-keystroke-tracking-ai-agents-2026), and [Reuters](https://www.reuters.com/legal/litigation/meta-pause-internal-mouse-tracking-tech-while-examining-data-security-issues-2026-06-22/) reported the programme was later paused while data security issues were examined. Measuring people by their tool usage sits uncomfortably close to surveillance, and employees notice the resemblance even when employers do not. Trust is expensive to rebuild once a workforce concludes that a productivity initiative is a monitoring initiative with better branding, and the reputational cost usually outlasts the policy that caused it. ## What is Hatch and why does it complicate the picture? Hatch is Meta's own agentic AI platform, still unreleased, which employees have had access to for several weeks according to WIRED. The company has been encouraging but not requiring staff to use it. Gizmodo notes the tool consumes more AI tokens than standard chatbots and coding assistants, which produces the genuinely odd situation the reporting describes. Meta has removed token counts from evaluations while simultaneously promoting an internal tool that drives token consumption up. That is less contradictory than it first appears, and the distinction is useful for any employee in a similar position. Removing a metric from a review is not the same as abandoning a strategic goal. The company still wants adoption. It has simply stopped making individual adoption a matter of personal assessment, which changes who carries the risk if the tool turns out not to help. Under the previous arrangement an employee who used the tool heavily and produced nothing was still meeting the measure, while an employee who tried it, judged it unhelpful and delivered anyway looked worse. Removing the metric puts that risk back where it belongs. Some of WIRED's sources said they remained wary of the new tool given the earlier keystroke episode, which is a reasonable position and one that is now easier to hold without consequence. That is arguably the most valuable part of the change for employees. Being able to decline a tool on considered grounds, without it appearing in an assessment, is what makes adoption evidence of usefulness rather than evidence of compliance. ## What does this signal for employers measuring AI use? Meta is a bellwether on workplace practice whether or not that is deserved, and this is the clearest signal yet that measuring AI usage directly has peaked as an idea. The organisations that copied the leaderboard approach are now looking at the company that originated it walking away. Expect the language in review templates elsewhere to move in the same direction over the next cycle, from usage of AI toward outcomes that may be supported by AI. If you sit in HR or people operations, this is the moment to make that change deliberately rather than to discover in six months that your review template is measuring something the market has moved past. "All three of the technologies mentioned, emotion inference, gait analysis and using facial recognition to discipline, have been built into employee screening and monitoring software, even though employees will probably never know they're there," Lacey Kaelani, chief executive of [Metaintro](https://www.metaintro.com), told [Reworked](https://www.reworked.co/talent-management/could-california-rein-in-the-ai-run-digital-workplace-or-will-the-bill-hit-a-wall/). The wider point holds here. Once a behaviour can be counted it tends to get counted, and employees are usually the last to learn which counters are running. The retreat from token metrics is welcome, but it is a retreat from one visible measure rather than a change in the underlying appetite to measure, which is why [AI workplace surveillance continues to reshape how work is done](https://www.metaintro.com/blog/ai-workplace-surveillance-breaking-how-we-work). ## How should you handle AI questions in your own review? Answer in outcomes, not in usage, because that is the direction the language is moving. Instead of describing which tools you adopted, describe what got faster, cheaper, more accurate or more reliable, and name the number that changed. If AI was part of how you achieved it, say so in one clause and move on. A reviewer who has been told outcomes can be supported by AI or other means is being invited to assess the outcome, and a candidate who leads with tool adoption is answering a question that is no longer being asked. Keep one or two concrete examples ready in which a tool changed how you worked and the result improved, since that format satisfies both the old framing and the new one without you having to guess which your reviewer holds. If your employer still measures usage directly, treat that as information about the organisation. Ask what the metric is meant to predict and what happens to it when it conflicts with quality. Reasonable managers can answer that. Where they cannot, you are working inside the arrangement Meta has just abandoned, and the sensible response is to keep your own record of outcomes so you are not solely represented by a dashboard. Write it as you go, with dates and figures, because reconstructing it during review season produces vaguer claims than recording it in the moment does. There has been real [worker backlash against forced AI adoption](https://www.metaintro.com/blog/worker-backlash-forced-ai-automation-workplace-2026), and it has generally worked where employees argued from results rather than from principle. ## What does this mean for your career right now? The durable lesson is about which metrics survive. Input measures like token counts, hours logged and tools adopted are easy to collect and easy to game, and they get abandoned once someone senior asks what they predict. Outcome measures are harder to collect and much harder to argue with. Building your own record around outcomes protects you through exactly this kind of policy reversal, because the evidence stays valid when the measurement regime changes underneath it. A record of what you delivered transfers to a new manager, a new template and a new employer. A record of how much you used a particular tool transfers nowhere at all. It is also worth being calm about the whiplash. Employers are experimenting in public with how to manage AI at work, and some of those experiments will be reversed within a year, as this one has been. That argues for treating any single mandate as temporary and for keeping your skills anchored to the work rather than to a tool. In a market where [job openings sit at 7.3 million against 5.1 million hires](https://www.bls.gov/news.release/jolts.nr0.htm), according to Bureau of Labor Statistics data for July 2026, the ability to demonstrate results is worth considerably more than the ability to demonstrate compliance. Performance review systems themselves are being rebuilt around this, which is the useful context in [how generative AI is changing workplace feedback](https://www.metaintro.com/blog/gen-ai-performance-reviews-2026-fix-workplace-feedback). ## Related Articles - [Meta mouse tracking and workers fighting back](https://www.metaintro.com/blog/meta-mouse-tracking-workers-fighting-back-2026) - [Block layoffs, AI mandates and employee backlash](https://www.metaintro.com/blog/jack-dorsey-block-layoffs-ai-mandates-employee-backlash-2026) - [Duolingo AI performance monitoring and workers pushing back](https://www.metaintro.com/blog/duolingo-ai-performance-monitoring-workers-push-back) - [Debunking performance review myths](https://www.metaintro.com/blog/debunking-performance-review-myths) - [The real threat of AI worker surveillance](https://www.metaintro.com/blog/ai-worker-surveillance-real-threat) - [Inside Meta's AI reorg and its career growth promises](https://www.metaintro.com/blog/meta-calls-ai-reorg-atrocious-career-growth-promises) - [What workers should know about performance improvement plans](https://www.metaintro.com/blog/pip-court-ruling-what-workers-should-know-2026) - [When AI gains become more work rather than time off](https://www.metaintro.com/blog/meta-cto-time-off-ai-gains-more-work-2026) ## People Also Asked ### Q: Is Meta still measuring employees on AI usage? A: No. WIRED reported that newly published performance review guidance no longer makes usage of AI a criterion for assessment, and The Information reported engineers were told AI adoption dashboards and token counts would no longer be part of evaluations. The updated guidance says desired outcomes can be supported by AI or other means, so AI remains encouraged without being the measure. ### Q: What is tokenmaxxing? A: The practice of maximising AI token consumption among staff, popularised by a March 2026 New York Times column by Kevin Roose which reported that AI enthusiastic workplaces including Meta were rewarding heavy users of AI tools and chastening those who did not use them, in some cases using leaderboards that turned consumption into a competition between colleagues. ### Q: What is Meta's Hatch tool? A: Hatch is Meta's own agentic AI platform, still unreleased publicly, which employees have had access to for several weeks according to WIRED. Staff are being encouraged but not required to use it. It reportedly consumes more tokens than standard chatbots and coding assistants, which sits oddly alongside the decision to stop counting tokens in reviews. Get the inside scoop on how employers actually evaluate people. [Metaintro](https://www.metaintro.com) matches you with roles based on the outcomes you have delivered rather than the tools on your profile. [Create your free profile](https://www.metaintro.com/signup) and see who is hiring for what you can do. ### Share this article [https://x.com/intent/tweet?text=Meta%20Stops%20Grading%20Engineers%20on%20AI%20Token%20Counts%20and%20What%20That%20Signals%20for%20Your%20Review&url=https%3A%2F%2Fwww.metaintro.com%2Fblog%2Fmeta-drops-ai-token-counts-performance-reviews-2026](https://x.com/intent/tweet?text=Meta%20Stops%20Grading%20Engineers%20on%20AI%20Token%20Counts%20and%20What%20That%20Signals%20for%20Your%20Review&url=https%3A%2F%2Fwww.metaintro.com%2Fblog%2Fmeta-drops-ai-token-counts-performance-reviews-2026)[http://www.facebook.com/sharer.php?u=https%3A%2F%2Fwww.metaintro.com%2Fblog%2Fmeta-drops-ai-token-counts-performance-reviews-2026](http://www.facebook.com/sharer.php?u=https%3A%2F%2Fwww.metaintro.com%2Fblog%2Fmeta-drops-ai-token-counts-performance-reviews-2026)[https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fwww.metaintro.com%2Fblog%2Fmeta-drops-ai-token-counts-performance-reviews-2026&title=Meta%20Stops%20Grading%20Engineers%20on%20AI%20Token%20Counts%20and%20What%20That%20Signals%20for%20Your%20Review](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fwww.metaintro.com%2Fblog%2Fmeta-drops-ai-token-counts-performance-reviews-2026&title=Meta%20Stops%20Grading%20Engineers%20on%20AI%20Token%20Counts%20and%20What%20That%20Signals%20for%20Your%20Review)[mailto:?subject=Meta%20Stops%20Grading%20Engineers%20on%20AI%20Token%20Counts%20and%20What%20That%20Signals%20for%20Your%20Review&body=https%3A%2F%2Fwww.metaintro.com%2Fblog%2Fmeta-drops-ai-token-counts-performance-reviews-2026](mailto:?subject=Meta%20Stops%20Grading%20Engineers%20on%20AI%20Token%20Counts%20and%20What%20That%20Signals%20for%20Your%20Review&body=https%3A%2F%2Fwww.metaintro.com%2Fblog%2Fmeta-drops-ai-token-counts-performance-reviews-2026) For job seekers ## Skip the job boards. Keep the career. Tell Metaintro your story once. It reads every posting, every day, and brings you only what fits. [Get Started Free](/signup)[Search matching jobs](/jobs/search) - MatchCompare live roles against your current evidence. - PositionTurn proof projects into role-specific applications. - ImproveUse market feedback to keep the skill plan current. [Return to navigation](#main-navigation)