Amazon memo discourages tokenmaxxing while keeping AI at the centre
A warning about metrics, not a ban
Amazon has reportedly told its software engineers that token consumption is a poor proxy for good engineering, in internal guidance that warns treating AI usage metrics as targets can increase costs and complexity even as the company says AI remains central to its work.
The guidance was circulated in September 2026 and, according to The Indian Express, posted to Amazon’s website on 27 October 2026. The Register and IT Pro also reported its contents. The reported guidance does not amount to a ban on AI. An Amazon spokesperson confirmed to The Register that the memo was intended to discourage tokenmaxxing, and IT Pro, citing CNBC, reported that Amazon said there was no cap on AI usage. Those are Amazon’s characterisations, relayed by secondary outlets, not independent confirmation of practice.
Tokenmaxxing describes the practice of using AI tools as heavily as possible, sometimes to climb internal usage rankings rather than to complete a task. Tokens are the units used to meter AI model usage, and they become a direct cost when providers charge by consumption rather than a flat subscription. The reported memo treats that activity as a problem, not an achievement.
According to the reports, the memo said token usage, AI tool adoption rates and code changes or reviews were poorly connected to what matters. Used as targets rather than measures, those numbers could drive higher costs and greater complexity. The memo instead pointed engineers towards value and long-term outcomes, and said work that does not benefit from current AI tooling should be treated as a useful finding rather than a personal shortcoming. It also said AI would remain central to software engineering and encouraged engineers to learn and apply it.
Why metrics can mislead
That distinction between a measure and a target is central. Token counts can show activity, but they do not show whether software is maintainable, reliable or useful to customers. When a number becomes a ranking or a performance goal, workers have an incentive to produce the number rather than the outcome it was meant to represent. The memo’s reported concern about complexity links the problem to maintenance: more AI-assisted code can mean more code to review and a larger maintenance burden. The Register also noted a tension in the evidence, reporting that Amazon had seen some correlation between AI input metrics and productivity on tokenmaxxing teams while still calling the metrics flawed.
Leaderboard, costs and code review
The memo did not arrive in isolation. The Indian Express, citing CNBC, reported that Amazon shut down an internal AI usage leaderboard in May 2026 after employees were found to be tokenmaxxing to climb its rankings. Yahoo Finance also reported that the leaderboard was discontinued. The sequence points to successive governance adjustments, but the available record does not establish a formal review that directly caused the late October guidance, a single company-wide policy or the leaderboard’s full rationale beyond its reported link to rankings.
Cost is the other part of the context, but the strongest figures are attributed. The Financial Times, as relayed by Yahoo Finance, reported that an internal Claude Sonnet deployment meant to match author details with product listings cost US$1.8 million, 860% above its original allocation, and went undetected for about five months. The same reporting described a financial auditing tool that exceeded its budget by US$541,000 and a logistics initiative with US$134,000 in unplanned costs. The figures remain unverified. Amazon’s response, as relayed by Yahoo Finance, said the company was experimenting, learning and improving, including on cost efficiency, and that isolated examples did not accurately capture how its teams use AI tools.
Earlier tooling guidance adds another layer, but it should not be folded into the cost story without care. Business Insider reported in February 2026 that Amazon employees could not use Anthropic’s Claude Code for production code or live products without formal approval, and that internal guidance for production work favoured Amazon’s Kiro assistant over non-approved third-party tools. Amazon told Business Insider there was no explicit ban on Claude Code, but that stricter requirements applied to tools used for production code, with a process for exceptions. It is not evidence that the October memo was driven by the same concerns, although both point to tighter governance of AI-assisted development.
Reliability questions have also been reported, with important qualifications. heise online reported that Amazon was introducing stricter controls for AI-generated code after outages in March 2026, with AI-assisted changes to be approved only after review by experienced personnel. That account rested on Financial Times reporting and unnamed internal sources. It also reported a December incident in which Amazon’s Kiro tool independently deleted and recreated a production environment, causing a 13-hour outage of an AWS cost calculator, while Amazon described a very small issue affecting a single service in parts of China. The scope is disputed. The reporting concerns cost, review and reliability, not a security breach.
What the record does not show
Read together, the record supports a narrower conclusion than a simple order to rein in AI use. Amazon is reported to have told engineers not to mistake consumption for value, not to stop using AI. The memo sits alongside a closed leaderboard, earlier tool restrictions and stricter code review, but the evidence does not establish a single cause, a company-wide retreat or a formal link between cost overruns and outages. Nor do the reported figures show that Amazon’s AI spending is out of control. They are attributed to unnamed sources and disputed in their broader significance by the company.
The memo’s logic points beyond Amazon. AI usage is easy to measure and hard to evaluate. Token consumption, adoption rates and code changes can be counted, but they can also be gamed or mistaken for value. Cost control requires visibility of spend, while software quality requires review that survives the pressure to generate more code faster. The reported guidance points towards outcomes, judgement and long-term value, but it is not proof of audited savings or a settled productivity model.
The available reports do not establish which teams the guidance covers, how exceptions are handled, or whether it has changed engineering practice. Amazon has disputed the broader characterisation of the reported cost overruns and the scope of the Kiro incident. Amazon also says it has not capped AI use and that the technology will remain central. The record therefore describes a company refining how it measures AI work, not a retreat from the technology; whether the guidance changes costs or code quality remains to be established.
A usage counter can show that a tool was used, but not whether the work it produced was sound. Software teams meet that gap whenever a number becomes a target, and security teams face a similar question when activity is mistaken for evidence of a control working. Cybersecurity fundamentals give that judgement a firmer base: understanding what a control is meant to do, how it is scoped and reviewed, and why a measure can be misleading. Readers who want to build that habit can explore The Hack Academy’s online cybersecurity training programme, which lists cybersecurity fundamentals among its subject areas and sets self-paced learning and practical exercises alongside the theory.