🔍 Read the full analysis: Amazon’s AI Bets And The Search For Better Engineering Metrics on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Amazon published engineering guidance on October 27 saying token counts and other AI-use metrics should not be targets for software engineers. The memo says such measures can help identify patterns but may raise costs and complexity when treated as goals. Amazon’s Nova models and its investment in Anthropic provide contrasting examples of the company’s AI strategy, though the financial and operational picture is still developing.
Amazon published revised guidance for its software engineers on October 27, warning that token counts, AI-tool adoption and other activity measures should not be used as performance targets. The memo says these figures can help teams spot patterns, but judging engineers by them risks encouraging higher costs and more complex systems instead of better results for customers, the business and colleagues.
The memo was written by Terk Terkowitz, a vice president and distinguished engineer in Amazon’s stores business, and presents a revised role guideline for software engineers. It says long-term outcomes are the right measure of work, while acknowledging that they can take time to assess. Some teams have used inputs such as token usage, AI adoption, code changes and reviews as quicker stand-ins. The memo calls those measures “deeply flawed” when used as targets, rather than as information to inform judgment.
The guidance does not say teams should stop using AI. It says these input measures currently “correlate well” with outcomes that matter, and describes teams that have adopted AI heavily as delivering more while some teams that have not adopted it are falling behind. Its distinction is between monitoring usage and rewarding usage for its own sake. The memo’s examples invoke Amazon leadership principles: tokenmaxxing conflicts with frugality, while measuring lines of code conflicts with inventing and simplifying.
It asks managers to examine both low and high levels of adoption: why some teams are not using AI effectively, and whether heavy users are producing better results. Engineers whose work does not benefit from current tools are asked to explain what they found, rather than treat that outcome as a failure. Leaders are also expected to use the tools in their own production work and describe tasks where they helped and where they did 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
From AI Activity to Engineering Outcomes
The guidance addresses a practical problem for companies adopting generative AI: usage is easier to count than value. Tokens, tool logins and code changes can show that a tool is being used, but do not by themselves establish that work improved. If organizations tie rewards or targets to those figures, employees may have reason to increase activity even when it does not help a customer or team.
Amazon’s position leaves room for measurement while warning against turning proxies into goals. That matters for engineers who may be asked to demonstrate AI adoption, and for managers deciding how to evaluate productivity. The memo says that a developer’s conclusion that AI does not help a particular task can be a useful finding. Its approach puts more weight on judgment and demonstrated outcomes than on a single company-wide adoption score.
The company’s broader AI choices add a business dimension. Amazon’s in-house Nova models and its investment in Anthropic represent different routes to building AI capabilities: developing models itself and backing an outside model maker. The available reporting suggests different results so far, but it does not settle the long-term commercial value of either route.
Products Worth Considering
Amazon Kindle Scribe Colorsoft 64GB (1st Generation)
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Jade
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black
As an affiliate, we earn on qualifying purchases.
Amazon’s Two AI Investment Paths
Amazon’s internal debate follows earlier reports of employees tracking AI activity. The Financial Times reported in May that some staff gave an internal agent tool, MeshClaw, unnecessary tasks to increase token counts. Those accounts were attributed to anonymous employees and should be treated as reported allegations, not as independently confirmed company-wide practice. At the end of May, Amazon confirmed to Business Insider that an employee-built leaderboard called KiroRank had been “deprecated.” The company said it was an informal tracker and had not been intended to encourage AI use for its own sake.
The new memo makes the response more formal by placing the measurement principle in an engineering role guideline. Separately, Business Insider reported in July that most of Amazon’s flagship Nova models had been moved to a “keep the lights on” status. AWS lifecycle information also lists Nova Canvas, Nova Reel and the first Nova Sonic as retired. Those details point to changes in parts of the Nova lineup, but do not establish that Amazon has abandoned its broader model-development work.
Amazon has also invested in Anthropic. The source reports about $18 billion invested and says an Amazon filing valued its holdings at roughly $190 billion as of June 30. Those are reported figures and the valuation is on paper; the value of an investment can change. Amazon is scheduled to report third-quarter results on October 29, 2026, which may provide further information on its finances and AI strategy.
““Tokenmaxxing is at odds with frugality.””
— Amazon engineering memo
What the Memo Leaves Open
The memo does not set out a replacement numerical score for engineering performance. It instead asks managers to assess whether tools improve outcomes and to examine examples at both ends of the adoption range. How teams will apply that judgment, and whether it will change formal performance reviews, is not specified in the available material.
The earlier reports about employees inflating token counts relied in part on anonymous accounts. Amazon confirmed that KiroRank was discontinued, but the available statements do not establish how widespread tokenmaxxing was or whether it affected formal evaluations. Nor do the reported Nova changes establish the full status of Amazon’s model program. The reported Anthropic valuation is a paper value, not a realized return, and its future value is uncertain.
Results Due October 29
Amazon is due to report its third-quarter results on October 29, 2026, one day after publishing the engineering memo. The results may offer more detail about the company’s financial position and investments, although the source material does not indicate that Amazon will provide a specific breakdown of AI returns.
For now, the clearest next step is how engineering leaders put the guidance into practice: whether they use AI activity data to investigate adoption patterns, while judging work by its effects on customers, the business and teams. The memo does not announce a new leaderboard or a replacement metric.
Key Questions
What did Amazon announce?
Amazon published revised software-engineering guidance on October 27, 2026, saying AI usage measures such as token counts should not become performance targets.
Is Amazon telling engineers to stop using AI?
No. The memo says AI adoption measures can correlate with outcomes and asks leaders to understand both low and high adoption. Its warning is against using activity as a goal without checking whether it improves results.
What does “tokenmaxxing” mean here?
It refers to increasing AI token usage for the sake of usage or a higher leaderboard position, rather than because a task benefits from the tool. Earlier reports described employees assigning unnecessary work to an internal agent to raise token counts.
What does the memo say engineers should do if AI does not help?
It says that this can be a finding rather than a failing. Engineers are asked to explain why the tools do not benefit their work and share that assessment with peers and managers.
What remains unclear about Amazon’s AI strategy?
The source material does not establish the full status or future plans for Amazon’s Nova models, how the company will apply the guidance in performance reviews, or whether its paper valuation of Anthropic will translate into a realized financial return.
Source: ThorstenMeyerAI.com
