Amazon tokenmaxxing is now official company vocabulary. On October 27, Amazon published a revised guide for its software engineers that calls token counts and similar AI usage numbers “deeply flawed” ways to judge productivity, according to Fortune. The memo first circulated inside the company in September. Its message is short: use AI where it helps, and measure the result, not the volume.
If your team has set AI usage targets, this memo is worth a careful read. Here is what it says, what it leaves out, and what to do with it.
What Amazon tokenmaxxing actually means
Tokenmaxxing is the habit of burning as many AI tokens as possible so your usage looks impressive. It is what happens when a company tracks consumption and people respond by consuming. Fortune reports that Amazon employees ran AI agents on pointless tasks to climb an unofficial internal leaderboard, and that Amazon shut the board down in May over unnecessary compute costs.
Amazon is not alone. Gizmodo reports that Uber’s engineers had reportedly spent the company’s entire annual AI budget on Claude Code by April, in about four months. Gizmodo also notes that Dictionary.com cited the trend when it picked “maxxing” as its 2026 word of the year. Those are reported claims, not figures either company published.
What the memo says
The author is Terk Terkowitz, a vice president and distinguished engineer in Amazon’s Worldwide Stores division. The main points, as reported by Fortune and Gizmodo:
- Counting tokens or lines of code is “deeply flawed” because neither reflects the quality of the work.
- Turning those numbers into targets raises costs: “tokenmaxxing is at odds with frugality.” It can also add complexity that cuts against Amazon’s “invent and simplify” principle.
- “If your work does not benefit from current AI tooling, that is a finding, not a failing.”
- Adoption metrics are not being dropped. The memo says they still “correlate well with the outputs that matter.”
- Managers should compare the 10th and 90th percentile of AI users on their teams: ask why the lightest users use little, and check whether the heaviest users actually deliver more.
The memo also frames the ideal engineer as a “Renaissance Developer” who blends AI with judgment and can explain where AI does not yet help.
Input metrics versus throughput metrics
eWeek’s breakdown sorts the flawed measures into two groups. Input metrics, such as token consumption and adoption of AI coding tools, show that someone is experimenting. They do not show that anyone is getting more done. Throughput metrics, such as code changes and reviews, can push people to produce unnecessary work when they become targets.
eWeek adds a useful caution about token counts in particular. It points to an AWS benchmark published October 21 that tested 12 models across two coding-agent frameworks. One model cost about $24 per task in the heavier framework versus $8 in the lighter one, without better scores. The test used tasks from a single repository and did not assess employees, so the finding is narrow: tool setup can change token use and cost with no matching gain in quality. A big token number may say more about the configuration than about the person.
What the memo does not say
Read it closely and it is less of a retreat than the headlines suggest. Amazon still sees a link between heavy AI use and delivery on what it calls “frontier teams,” its most active users, though the coverage does not include the data behind that claim. A correlation like that could also mean strong teams adopt tools faster, not that the tools made them strong. Amazon says it will keep looking for “better metrics and smarter goals,” which is an admission that it does not have them yet.
It also does not ban tracking. Amazon has tracked token usage internally for cost management, and the memo keeps that as one signal among several.
How to avoid tokenmaxxing on your own team
You do not need Amazon’s scale to run into this. Any team told to “use AI more” will find a way to show it. A few practical steps:
- Measure outcomes first. Delivery time, rework, review effort and customer impact beat raw usage. Our look at how fast coding agent adoption is growing shows why adoption alone says little.
- Pair usage data with quality data. A heavy user with high rework is a different story from a heavy user who ships clean work.
- Check the setup before comparing people. Model choice and agent framework change token use. Cheaper models such as the ones in our Claude Haiku 5.5 price cut breakdown shift the cost picture too.
- Make “AI didn’t help here” a safe answer. This is the memo’s most useful idea. If people fear looking behind, they will inflate usage instead of reporting honestly.
- Watch spend and limits together. Providers are already tightening quotas, as we covered in AI usage limits tightening this month, so wasted tokens now have a real ceiling.
The takeaway
The Amazon tokenmaxxing memo is a sign that the first wave of “use more AI” mandates is meeting reality. Usage is easy to count and easy to fake. Results are harder to measure, but they are the only number that matters. For a wider view of where enterprise adoption stands, see our post on enterprise AI agent adoption in 2026.
Sharing Is Caring