What happened: Some Amazon staff chased high token counts (the chunks of text an AI reads and writes) because internal leaderboards rewarded usage over results. In May, Amazon shut one down. On Oct 27, a public memo went further.
CNBC
How it played out
Goal: 80%+ of developers using AI weekly.
tomshardware
Some staff reportedly used an in-house agent platform to run needless tasks and boost their numbers.
tomshardware
Every extra token is real compute cost.
What the memo says
Token usage, tool adoption, and code-change counts are “deeply flawed” as targets.
CNBC
aboutamazon
Judge AI by long-term outcomes: customers, business, team.
If AI doesn’t fit your task, say so and explain why. That’s judgment, not failure.
Used with care, usage metrics still correlate with output.
Why devs should care
Measure shipped value, not prompts sent.
AI writes fast; you still own quality, so review carefully.
Spec-first prompts (like Kiro’s requirements, design, tasks flow) cut ambiguity.
Tags: #ai #softwareengineering #productivity #career
Media
Image: Amazon News header graphic (official page image). https://assets.aboutamazon.com/dims4/default/7e16f44/2147483647/strip/true/crop/2000x1000+0+63/resize/1200x600!/quality/90/?url=https%3A%2F%2Fassets.aboutamazon.com%2F52%2Fa8%2F3c8ee3984dd9a2548f89d35d787c%2Famazon-logo-hero-images-2024-4.jpg
Video: Werner Vogels’ re:Invent 2025 closing keynote (AWS Events). It covers the “Renaissance Developer” idea and a Kiro spec-driven coding demo. https://www.youtube.com/watch?v=3Y1G9najGiI
References & Credits
CNBC: https://www.cnbc.com/2026/10/27/amazon-ai-coding-tokenmaxxing.html
Amazon News, “An update from Amazon engineering leadership…” (Terk Terkowitz): https://www.aboutamazon.com/news/company-news/amazon-software-engineer-role-guideline-ai
Tom’s Hardware, Luke James (reporting the Financial Times): https://www.tomshardware.com/tech-industry/big-tech/big-tech-has-a-tokenmaxxing-habit
AWS Events, re:Invent 2025 keynote: https://www.youtube.com/watch?v=3Y1G9najGiI
Top comments (1)
The useful distinction buried in this: gaming a usage leaderboard and measuring token cost are opposite habits that ended up with the same name. "Tokens sent" as a target is pure vanity - you can inflate it with busywork, which is exactly what got shut down. But most teams overcorrect and measure nothing, then get surprised by the bill. The number that actually survives is "tokens re-paid" - how much of each turn is context being re-sent because something invalidated the cache, not new work. That tracks real compute and can't be gamed by running needless tasks: busywork adds turns, each turn re-sends the prefix, and that shows up as cost, not as a win. So the honest split is: measure shipped value for the human, and cache-read vs cache-write for the machine. The second is where the waste actually hides, and it's completely invisible on a "tokens used" leaderboard.