ai
The AI race at work is no longer about tokenmaxxing
Meta, Microsoft, Uber and Amazon have backed away from token counts as a measure of engineering work. Here is what they track now, and what that means for developers.
On 7 October, Microsoft proposed moving more AI agent work off the cloud and onto developers' own machines, R&D World reported. It is the latest sign that large employers no longer treat heavy AI use as a goal in itself. Six months ago, leaderboards ranking engineers by token use were running inside several of them.
What changed
Tokens are the unit AI providers bill by. The Associated Press describes them as small pieces of text a model reads or writes, each about three quarters of a word. Tokenmaxxing means using as many as possible, on the theory that heavy use marks a productive employee.
Four companies, when token use was the target and now
| Company | When tokens were the target | What it does now |
|---|---|---|
| Meta | "Claudeonomics" leaderboard covering 85,000+ staff | Token counts not used to evaluate impact (September) |
| Microsoft | Internal token leaderboard since January | Token budgets per division, cheaper default models (August) |
| Uber | Used its 2026 token budget in four months | Prompt caching, a new default model, per-hour cost visibility (August) |
| Amazon | "Kirorank" leaderboard for its Kiro coding tool | Leaderboard shut down in late May |
Sources: The Pragmatic Engineer, Fortune and HCAMag for the middle column; InfoWorld, ITPro, a later Fortune piece and HCAMag for the last.
Every row moves the same way: from counting use to limiting or pricing it.
How token counts became a target
In March, Nvidia chief executive Jensen Huang said he would be "deeply alarmed" if an engineer paid $500,000 a year spent less than $250,000 on tokens, The Decoder reported. By April, leaderboards had spread.
Once the number was visible, people optimized for it. One Microsoft engineer told The Pragmatic Engineer they asked AI about documentation they could have read directly, and built prototypes they planned to discard. At Amazon, some staff assigned agents to needless tasks to climb the ranking, according to Financial Times reporting summarized by HCAMag.
Not a retreat from AI
Uber's chief technology officer, Praveen Neppalli Naga, wrote in August that the company had quadrupled the number of employees using frontier AI tools while lowering its cost per token. Uber treated efficiency, he wrote, "as an engineering problem rather than a budget problem". Fortune, which reported the post, noted that he did not say whether total usage had gone up or down.
Microsoft took a similar line. R&D World quotes an August message to engineers, first reported by 404 Media, as "tokenmaxxing is not what we are optimizing for". This week's move goes further: its MAI-Code-1.1-Flash coding model will run on local hardware, which removes the per-request charge.
Two caveats apply. Much of this evidence is internal memos reported secondhand. And no company named here has published data showing that its new measures track value better than token counts did.
Why it matters
For working developers, the signal has flipped. Cost per useful result now carries the weight that token volume did. The techniques being reported are ordinary engineering: cache prompts, choose the smallest model that does the job, and know what a session costs. As Bain consultant Jue Wang told the Associated Press, "Not everything needs a Claude Opus 4.6."
The takeaway
Token counts measured an input, and employers found that input easy to inflate and expensive to reward. Whether the replacement measures capture value any better is still open. The thing to watch is the first company that publishes outcome data, not another memo.
Sources
- R&D World — After tokenmaxxing backlash, Microsoft moves more agent work onto the desk (7 October 2026)
- InfoWorld — Meta minimizes role of token maxing in employee evaluations (4 September 2026)
- ITPro — Microsoft has joined the growing list of companies cracking down on 'tokenmaxxing' (10 August 2026)
- Fortune — Uber CTO says, 'We're coming to the end of the so-called tokenmaxxing era' (7 August 2026)
- Associated Press via WTOP — AI 'tokenmaxxing' fades as workplaces look to cut tech spending (28 July 2026)
- HCAMag — Amazon shuts down AI leaderboard after 'tokenmaxxing' (1 June 2026)
- Fortune — Tokenmaxxing is over (28 May 2026)
- The Pragmatic Engineer — The Pulse: 'Tokenmaxxing' as a weird new trend (23 April 2026)
- The Decoder — Jensen Huang on $250K token budgets (21 March 2026)
More articles
How grill, spec, review, retro works: four gates for building with coding agents
Grill before you spec, spec before you build, review before you ship, retro after. What each gate is for, what it should leave behind, and when to skip it.
How time-boxing worker agents works, explained with a four-quarter rule
One developer gives each coding agent an hour, split into four quarters: commit by minute 30 or be stopped, and hand off in the last 15. Here is why that works, and where it breaks.