🔍 Read the full analysis: Amazon’s AI Investments Prompt A Rethink Of Tokenmaxxing on ThorstenMeyerAI.com
Get the latest gadgets 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 October 27 saying token use and other AI activity metrics should not be treated as performance targets. The memo says those measures can help identify patterns but may raise costs and complexity when optimized for their own sake. The guidance lands amid reported changes to Amazon’s Nova models and a large, paper-valued investment in Anthropic.
Amazon published revised guidance for its software engineers on October 27, warning that token counts and other measures of AI activity should not be used as performance targets. The memo says those inputs may correlate with useful results today, but treating them as goals can increase costs and system complexity without showing whether customers or the business benefit.
The memo was written by Terk Terkowitz, a vice president and distinguished engineer in Amazon’s stores business, and describes revised role guidance developed by the company’s senior engineers. It defines engineering success around impact on customers, the business and the team, rather than how much AI a worker uses. The authors say detailed AI-specific job rules could quickly become outdated as tools change.
Amazon identifies token usage, AI-tool adoption, code changes and code reviews as possible input measures. Those measures can be informative, the memo says, and currently “correlate well” with outcomes that matter. But the document calls them “deeply flawed metrics” when used as targets instead of measures: optimizing for them can drive up costs and make systems more complex. It specifically says “tokenmaxxing is at odds with frugality” and measuring lines of code conflicts with the principle to “invent and simplify.”
The guidance asks managers to examine both low and high AI adoption. Looking at teams around the 10th percentile can help explain why some are not using the tools effectively; examining those around the 90th percentile can test whether heavy use is producing more value. Engineers are also asked to document when current AI tools do not help, while leaders are told to use the tools in their own work and explain where they did and did not help.
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
Amazon Shifts AI Performance Measures
The policy matters because it separates AI activity from engineering performance inside one of the largest technology companies. Token counts can show that a tool is being used, but they do not by themselves establish that work improved. If employees are rewarded for activity rather than results, they may spend time generating unnecessary output or add complexity in pursuit of a visible number.
Amazon’s approach is not a rejection of AI adoption. The memo says teams using AI more are currently delivering more, while teams that do not adopt it are falling behind. Its stated position is that adoption data can help leaders find where tools work, but should not become a quota or individual score. The distinction is relevant to workers and employers weighing how to judge productivity as AI coding tools become more common.
The company’s wider AI portfolio gives the memo added significance, though it does not prove that the guidance was prompted by results from any specific model. Amazon’s reported retrenchment from some Nova models and its substantial Anthropic investment illustrate why usage and investment alone are incomplete scorecards: the practical question is whether a product or tool delivers outcomes at a cost the company and its customers can justify.
Products Worth Considering
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black
Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI-ARC, Black
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Jade
As an affiliate, we earn on qualifying purchases.
From Internal Leaderboards to Guidance
The memo follows reports of internal efforts to track AI use. In May, the Financial Times reported that some Amazon employees gave the internal agent tool MeshClaw unnecessary tasks to raise token counts. The report also described a weekly AI-use target for developers and internal leaderboards. Those accounts were based on reporting and anonymous employee claims; they are not independently established by the October memo.
Amazon did confirm to Business Insider at the end of May that an informal employee-built leaderboard called KiroRank had been deprecated. The company said it had not been intended to encourage 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 new role guidance makes the outcome-focused principle part of official engineering expectations.
Amazon’s model and investment moves provide separate corporate context. Business Insider reported in July that most flagship Nova models had been moved to a “keep the lights on” status. Amazon’s lifecycle information lists Nova Canvas, Nova Reel and the first Nova Sonic as retired. Separately, the source material says Amazon invested about $18 billion in Anthropic and reported the holding at about $190 billion in a June 30 filing. That figure is a reported paper valuation, not cash proceeds, and the source says second-quarter gains on the stake exceeded Amazon’s operating income for that quarter.
““Tokenmaxxing is at odds with frugality.””
— Amazon engineering memo
What the Memo Does Not Resolve
The guidance does not set out a single replacement formula for measuring AI’s effect on productivity. It does not quantify how much time, cost or customer value AI has produced across Amazon, nor does it specify how managers should weigh different outcomes across teams. The memo presents judgment and examination of high- and low-adoption groups as the response, rather than a new leaderboard.
The reports about employees inflating token counts rely in part on anonymous accounts and should be treated as reported allegations, not as findings confirmed by the memo. The company’s earlier description of KiroRank as informal is on the record, but does not establish the full scale of tokenmaxxing. The status of Amazon’s Nova products and the valuation of Anthropic are also separate developments; the available material does not show that either directly caused the new engineering guidance.
Amazon’s third-quarter results were due on October 29. At the time of the source report, those figures had not yet been released, so the memo should not be read as a statement of quarterly financial performance or as evidence that Amazon has settled the return on its AI spending.
Amazon’s Results and Policy in Practice
The next scheduled company milestone is Amazon’s third-quarter earnings report on October 29. That release may provide financial information relevant to the company’s broader business, but it will not by itself show how much value particular AI tools deliver unless Amazon supplies specific figures and context.
For the engineering policy, the next test is how teams apply the guidance: whether managers use adoption figures to investigate barriers and benefits, rather than to rank workers, and whether engineers’ documented examples influence tool choices. Amazon’s memo says leaders should try AI on their own production work and explain cases where it helps and where it does not. The company has not provided a timetable or public measures for evaluating the policy’s effect.
Key Questions
What is tokenmaxxing?
Tokenmaxxing means encouraging or carrying out AI tasks to increase the amount of text an AI model processes, rather than because the work is useful. The source report says some Amazon engineers used an internal tool for unnecessary tasks to raise token counts; that account was reported, not confirmed in the memo.
Is Amazon telling engineers to stop using AI?
No. The memo says AI adoption can correlate with useful outcomes and that teams not adopting the tools may be falling behind. Its stated warning is against treating usage itself as a target, not against using AI where it improves work.
What does Amazon say managers should measure instead?
The memo centers on customer, business and team impact, while acknowledging those outcomes can take time to measure. It asks managers to study both low- and high-adoption teams and judge whether heavier use leads to better results.
Did Amazon confirm that employees inflated token counts?
The October memo does not confirm the reported behavior. The Financial Times reported in May that some employees assigned unnecessary tasks to an internal agent to increase token use. Amazon separately confirmed that the KiroRank leaderboard had been deprecated and described it as an informal tracker.
What happens next?
Amazon’s third-quarter results were scheduled for October 29, 2026. The longer-term question is how the company applies the new guidance and whether it reports evidence that AI use is improving outcomes rather than simply increasing activity.
Source: ThorstenMeyerAI.com
