🔍 Read the full analysis: Beyond Tokenmaxxing: How Amazon’s AI Bets Suggest Measuring Results 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 an engineering guideline on October 27 saying AI usage measures such as token counts should not be used as targets for judging engineers. The memo says those measures can correlate with useful output, but the test should be whether work improves for customers, the business and teams. Amazon’s Nova model decisions and its investment in Anthropic offer contrasting company-level examples, though they do not by themselves prove which AI strategy will produce better returns.
Amazon published new guidance for its software engineers on October 27, warning that token counts and other measures of AI use should not become performance targets. The guideline says engineers and managers should judge AI by its impact on customers, the business and teams—a distinction that matters as companies try to determine whether growing AI use is producing results rather than activity.
The memo was written by Terk Terkowitz, a vice president and distinguished engineer in Amazon’s stores business, and sets out a revised role guideline for software engineers. It says long-term outcomes are the measures that matter, while recognizing that they can be slow and difficult to assess. Some teams have therefore relied on indicators such as token usage, AI-tool adoption, code changes and reviews.
Amazon’s guidance calls those indicators “deeply flawed” as targets. Used to measure activity, they can be informative; turned into goals, the memo warns, they can raise costs and add complexity without showing whether the work is valuable. The memo says these measures currently “correlate well” with outputs that matter, so it does not recommend ignoring them. Instead, it argues they should be treated as signals rather than a score for individuals or teams.
The guideline asks leaders to examine both low and high AI adoption. Looking at teams near the 10th percentile can help identify why tools are not working for them; examining those near the 90th percentile can show whether heavy use is translating into better delivery. It also asks leaders to use AI on their own production work and describe a task where it helped and one where it 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 Measurable Outcomes
The memo addresses a problem facing employers adopting generative AI: usage is easy to count, but value is harder to establish. A token total or the number of AI-assisted code changes can show that a tool was used. It does not, by itself, show that customers received a better product, that work was completed faster, or that costs fell. If workers are rewarded for usage alone, they may have reason to produce more activity rather than better results.
Amazon’s approach offers a practical distinction for companies setting AI targets. Usage data can help locate teams that need support or identify practices worth examining, but it should not substitute for outcome measures. The memo also makes room for engineers to report that a tool is not helping a particular task: not using AI can be a considered judgment, rather than evidence of poor performance.
The company’s broader AI investments provide a reason to examine that distinction, but they should not be mistaken for a controlled comparison. Amazon’s own model portfolio and its investment in Anthropic reflect different strategic choices; reported financial values and product decisions can help describe those choices, but do not alone establish the eventual return from either.
Products Worth Considering
DJI Osmo Pocket 3 Filter Bundle – Vlogging Camera with 1 CMOS Sensor and 4K/120fps
Jackery Solar Generator 1000 v2 with 200W Solar Panel
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
As an affiliate, we earn on qualifying purchases.
Amazon’s Models and Anthropic Bet
The guidance follows earlier reports of internal efforts to track AI use. In May, the Financial Times reported that some Amazon staff gave an internal agent tool, MeshClaw, unnecessary tasks to increase token counts. That account was based on reported employee descriptions, not a public company finding. 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 was never intended to promote AI use for its own sake. Senior vice president Dave Treadwell also told staff not to use AI simply to use it.
Amazon’s decisions about its own models add context to the memo’s emphasis on outcomes. Business Insider reported in July that most flagship Nova models were moved to “keep the lights on”. AWS’s lifecycle page lists Nova Canvas, Nova Reel and the first Nova Sonic as retired. These are product and lifecycle developments, not a direct measure of the models’ full commercial performance.
Amazon also invested in Anthropic, whose models it does not control. The source material says Amazon invested about $18 billion and that a June 30 filing valued its holdings at about $190 billion. Those figures describe an investment and a reported paper value; they do not establish realized proceeds or the long-term returns from the partnership. Amazon is scheduled to report third-quarter results on October 29, the day after the memo’s publication.
““Tokenmaxxing is at odds with frugality.””
— Amazon’s engineering guideline
What the Guidance Cannot Yet Show
The memo does not provide a company-wide measure of how much AI has improved engineering productivity, customer outcomes or costs. It says usage indicators currently correlate with important outputs, but does not publish the data behind that assessment or specify how teams should measure impact across different kinds of work. The size and durability of any productivity gains remain unclear.
The investment figures also need careful reading: the reported value of Amazon’s Anthropic holdings is a paper valuation, not proof of cash returns. The Nova lifecycle decisions do not, by themselves, establish the commercial performance of every Nova product. The source material provides no third-quarter results, and Amazon’s scheduled October 29 report had not yet taken place at the time of the October 27 memo.
Amazon’s Results and Measurement
Amazon’s next scheduled financial milestone is its third-quarter earnings report on October 29, 2026. That report may add information about the company’s finances, but the source material does not indicate that it will include a detailed accounting of AI’s effect on engineering work or returns from individual model investments.
For the engineering guideline, the next test is how managers put it into practice: whether they use adoption data to investigate differences between teams, and whether they assess AI against meaningful work and customer outcomes rather than raw activity. The memo sets a standard, not a published results scorecard; evidence of how consistently Amazon applies it is not yet available.
Key Questions
What did Amazon announce on October 27?
Amazon published a revised engineering role guideline saying AI use should be judged by its effect on outcomes, not by token counts or other usage measures treated as targets.
Is Amazon telling engineers to stop using AI?
No. The memo says AI adoption can correlate with useful outputs. Its point is that usage alone is not a measure of success, and engineers should explain when a tool does not help their work.
Which AI metrics does the memo discuss?
It names token usage, AI-tool adoption, code changes and reviews. Amazon says such indicators can be useful signals, but are flawed when used as performance targets disconnected from results.
What is tokenmaxxing?
In the memo’s context, tokenmaxxing means trying to increase AI token usage for its own sake, including through unnecessary tasks. Amazon’s guideline says that behavior conflicts with frugality and outcome-focused work.
What remains unknown about Amazon’s AI results?
The memo does not publish a company-wide estimate of AI’s effects on productivity, costs or customer outcomes. The long-term returns from Amazon’s model choices and its Anthropic investment also cannot be established from the figures cited alone.
Source: ThorstenMeyerAI.com
