🔍 Read the full analysis: Why Amazon Is Telling Engineers To Stop Tokenmaxxing—and What Its AI Bets Measure on ThorstenMeyerAI.com
Get movie nights delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Amazon published engineering guidance on Oct. 27 saying software engineers should be judged on outcomes, not token use or other AI activity metrics. The memo says those measures can still help identify adoption patterns, but turning them into targets risks higher costs and unnecessary complexity.
Amazon published new guidance for software engineers on October 27, warning that chasing AI token counts and other usage metrics can raise costs and complexity without proving that work is improving. The memo says engineers should be evaluated on their impact on customers, the business and their teams, while AI activity data remains a tool for examining adoption—not a scorecard.
The guidance was written by the company’s most senior engineers and introduced by Terk Terkowitz, a vice president and distinguished engineer in Amazon’s stores business. It revises the role guideline for software engineers across the company. The memo says the role should be defined around durable responsibilities rather than specific AI requirements, since technology and tools change quickly.
The memo identifies token use, AI-tool adoption, code changes and code reviews as inputs that can be mistaken for meaningful results. Used as targets, they can encourage activity that does not help customers and can add cost or system complexity. The memo gives tokenmaxxing—using an AI tool for unnecessary tasks to increase token counts—as an example that conflicts with Amazon’s stated principle of frugality.
Its message is not to stop using AI. The memo says these metrics currently “correlate well” with outcomes that matter, and that teams adopting AI are delivering more. It recommends looking at both low and high adoption: leaders should investigate why some teams use the tools less, and check whether heavy users are producing better results. The guidance also asks leaders to use the tools on their own production work and explain where they help and where they do not.
Six months at Amazon
From token leaderboards to an outcomes memo
Amazon spent 2026 learning the same lesson twice: once with how its engineers use AI, once with its own AI models.
- May 2026Token leaderboards backfireFinancial Times (via The Decoder): staff give an internal agent tool pointless tasks to inflate token counts. Reported, anonymous sources.
- Late May 2026KiroRank “deprecated”Amazon spokesperson to Business Insider. SVP Dave Treadwell: “Please don’t use AI just for the sake of using AI.”
- 24 Jul 2026San Francisco AGI Lab site closesConfirmed by Amazon to GeekWire. Frontier model research continues under Pieter Abbeel.
- 28 Jul 2026Most flagship Nova models wound downBusiness Insider, people familiar: Premier, Omni, Reel, Canvas to “keep the lights on”. Amazon: models “remain one of the most important things we’re working on”.
- 14 / 30 Sep 2026Nova Sonic v1, Nova Reel, Nova Canvas reach end of lifeAWS Bedrock model lifecycle page. Requests to them now fail.
- 26 Oct 2026Anthropic files its S-1Amazon notes: $6.2 billion principal, 12% ownership cap that Amazon can waive. Anthropic commits at least $110 billion to Amazon through April 2036.
- 27 Oct 2026The memo: outcomes, not tokenmaxxingUsage and throughput metrics are “deeply flawed” as targets. “Tokenmaxxing is at odds with frugality.”
- 29 Oct 2026Amazon’s third-quarter resultsAnnounced by Amazon. Not yet published when this article went live.
Sources: Amazon engineering memo (27 Oct 2026); Business Insider (29 May and 28 Jul 2026); GeekWire (24 Jul 2026); The Decoder (12 May 2026, summarising the Financial Times); AWS Bedrock model lifecycle (read 28 Oct 2026); Anthropic S-1 (26 Oct 2026).
Powered by Thorsten Meyer AI
Why Amazon Is Changing Its AI Scorecard
The policy addresses a problem facing companies measuring AI adoption: usage is easy to count, but value is harder to establish. A token tally can show that a tool is being used, but by itself it cannot show whether a task was completed faster, whether the result was better, or whether the work benefited customers. Making usage a target can reward the visible activity rather than its value.
For Amazon engineers, the distinction affects how managers interpret both low and high adoption. An engineer who finds that current AI tools do not help with a particular task is not necessarily failing, the memo says; documenting that finding can help colleagues and managers understand where the tools fit. For the company, the policy signals that adoption remains important, but it is not meant to substitute for evidence of impact.
Popular right now
Dyson V8 Cyclone Cordless Vacuum
Vitamix 5200 Professional-Grade Blender, 64 oz. Container, Black
Ninja Flip Toaster Oven & Air Fryer, 8-in-1, Stainless Steel, SP151
As an affiliate, we earn on qualifying purchases.
From Internal Leaderboards to Formal Guidance
The memo follows earlier reporting about internal efforts to track AI use. In May, the Financial Times reported that some Amazon employees gave an internal agent tool, MeshClaw, unnecessary tasks to raise token counts on leaderboards. That account was based on employee reports and is not the same as a company-confirmed finding. The Financial Times also reported that Amazon had set a target for more than 80% of developers to use AI weekly.
Amazon did confirm to Business Insider at the end of May that a leaderboard called KiroRank had been “deprecated.” The company described it as an informal tracker created by a group of employees and said it was not intended to promote AI use for its own sake. Senior vice president Dave Treadwell also told staff, “Please don’t use AI just for the sake of using AI.” The October guidance turns that principle into formal engineering role expectations.
The source article also compares the memo with Amazon’s broader AI investments, including its Nova models and its investment in Anthropic. It cites Business Insider reporting that most flagship Nova models were moved to maintenance, and Amazon lifecycle information showing some Nova products retired. It reports that Amazon’s Anthropic holding was valued at about $190 billion in a June 30 filing after roughly $18 billion in investment. Those figures provide a separate measure of the company’s AI exposure; they do not establish why Amazon made its engineering policy change.
““Tokenmaxxing is at odds with frugality.””
— Amazon engineering memo
What the Memo Does Not Resolve
The memo does not specify a replacement numerical score for measuring AI’s impact, or describe a standard method for comparing productivity across teams. It calls for leaders to use judgment and examine adoption at both ends of the range, but the details of how managers will apply that guidance are not clear.
It is also uncertain how consistently the revised role guideline will change performance reviews or team-level targets. The memo says usage metrics can correlate with important outputs while warning against making them goals; it does not provide data quantifying that relationship or the costs linked to tokenmaxxing. Reports about employees deliberately inflating usage remain attributed accounts, not confirmed company findings.
Amazon’s Results and Policy Rollout
Amazon is scheduled to report its third-quarter results on October 29, two days after the memo’s publication. Those results may provide financial context for the company’s AI investments, but they will not by themselves show how the new engineering guidance is being applied internally.
The next evidence to watch is whether Amazon’s teams change how they set AI adoption goals and evaluate engineering work. The memo calls for leaders to test the tools on production tasks, record where they help and where they do not, and share those findings. The company has not yet detailed a companywide reporting framework or announced further changes to its AI performance measures.
Key Questions
What does “tokenmaxxing” mean at Amazon?
Tokenmaxxing means using an AI tool for unnecessary tasks to increase the number of tokens it processes, often to improve a usage ranking. Amazon’s memo says that behavior conflicts with frugality.
Is Amazon telling engineers to stop using AI?
No. The memo says AI adoption can be useful and that usage measures currently correlate with important outputs. Its warning is against treating usage itself as a performance goal.
What should managers measure instead?
The guidance centers on impact on customers, the business and the team. It asks leaders to investigate both low and high AI adoption and determine whether the tools are improving the work.
Did Amazon confirm that employees inflated AI usage?
No company confirmation of the reported behavior is provided in the source material. The Financial Times reported employee accounts about unnecessary tasks and token counts; Amazon separately confirmed that its KiroRank leaderboard had been deprecated and described it as an informal tracker.
When will Amazon report its next quarterly results?
Amazon is scheduled to report third-quarter results on October 29, 2026, according to the source material. Those results may offer financial context for its AI business, but are separate from the engineering memo.
Source: ThorstenMeyerAI.com
