Amazon Unplugs Kirorank After Staff Game the Token Score

Amazon shut Kirorank in late May 2026 after staff gamed an unofficial Kiro usage board through tokenmaxxing. The ranking lasted a few weeks. Senior vice president Dave Treadwell told employees not to use AI just to use AI, and Amazon said the beta dashboard was never an approved tool.

The company still wants more than 80 percent of developers on AI each week. It stopped keeping that score on a public internal board.

Kirorank Was Never an Official Amazon Tool

Employees were told in late May 2026 that Kirorank, which scored users of Amazon’s Kiro developer platform on AI activity, had been taken offline. People familiar with the matter said some workers assigned AI agents, bots that can act for a user, to needless jobs so they could climb the ranks.

Treadwell told staff the board had been built with “good intentions.” He also said the extra token burn, the units of data a model processes, had raised costs. “Please don’t use AI just for the sake of using AI,” he said. “Use AI to help you solve customer problems, to help you solve business problems, to innovate.”

One of the internal dashboards, called Kirorank, was recently created by a group of employees who wanted to drive awareness for how AI can accelerate work, and was never intended to promote the use of AI for usage’s sake. The beta dashboard was not a formal or approved tool, and has since been deprecated. We’re focused on AI adoption and sharing best practices to celebrate innovation and operational efficiency gains across the company, and we’re proud of the way our teams are embracing this technology.

Amazon spokesperson, email statement

Amazon said it still measures token use to understand cost and efficiency, and that it does not treat those counts as a measure of developer output. The board that made the counts visible, and ranked people on them, is the piece that came down.

The 80% Weekly Target Set the Incentive

Kirorank did not appear in a vacuum. Amazon had already set a target for more than 80 percent of developers to use AI each week, and it wanted that work on Kiro, the in-house coding assistant launched in July 2025.

A November 2025 memo signed by Peter DeSantis, who leads AWS utility computing, and Treadwell, who oversees eCommerce Foundation, named Kiro as Amazon’s “recommended AI-native development tool.” The memo said, “We do not plan to support additional third-party AI development tools.” OpenAI’s Codex was marked “Do Not Use” after a six-month review. Anthropic’s Claude Code briefly carried the same tag before that label was reversed.

A spokesperson said about 70 percent of Amazon’s software engineers used Kiro at least once in January 2026. That figure is a trial count, not the weekly target. About 1,500 staff had backed formal use of Claude Code on an internal forum, with one writing that a tool which cannot keep pace with rivals “offers no real innovation.”

THE KIRO ADOPTION PRESSURE

• Weekly target: More than 80 percent of developers were expected to use AI each week.
• January 2026: About 70 percent of software engineers used Kiro at least once.
• Board life: Kirorank existed for a few weeks and was never a formal tool.
• Official line: Tokens are tracked for cost and efficiency, not as a productivity score.

Rank people on a number the company already wants to rise, and the number will rise. Staff who feared looking idle on AI had a simple way to look busy: spend tokens.

FROM KIRO’S LAUNCH TO THE BOARD’S END

1. July 2025: Amazon launches Kiro as its AI-native coding tool.
2. November 2025: DeSantis and Treadwell circulate the memo that names Kiro the recommended tool and cools support for extra third-party coding products.
3. January 2026: About 70 percent of software engineers have used Kiro at least once.
4. April 2026: Meta’s Claudeonomics board is exposed and taken down; Uber’s Claude Code budget for the year is already gone.
5. Late May 2026: Amazon takes Kirorank offline and Treadwell tells staff not to use AI just to use AI.

The sequence is the point of the joke Amazon did not mean to tell. First came the quota. Then came the scoreboard. Then came the bill.

Agents Got Tasks That Solved Nothing

Tokenmaxxing is the habit of driving up token counts so a person, or a team, looks like a heavy AI user. On Kirorank, that meant spinning up agents and giving them menial work that did not need doing. The board counted activity. It did not ask whether a customer problem moved.

HOW THE GAME WORKED

• The score: Kirorank ranked people on Kiro activity, which showed up as tokens.
• The move: Staff created agents and pointed them at extra tasks so the count would climb.
• The bill: Each of those loops still consumed model time the company paid for.
• The tell: Amazon’s own note was to stop using AI for the sake of using AI.

That is not a mysterious failure. It is what happens when a measure becomes a target. A usage board is a perfect tracker of usage. It is a poor tracker of shipped work. People who wanted to look like adopters did not have to ship. They had to burn.

The louder problem is that Amazon pays the meter. A private contest on someone else’s API is a rounding error. A contest on the company’s own coding agents is a line item. Kirorank made that line item a sport.

Meta’s Token Legend Board Died in Public

Amazon was late to this particular mess. In April 2026 a Meta employee’s unofficial board, Claudeonomics, ranked more than 85,000 staff by token use and posted the top 250. It handed out titles such as Token Legend and Cache Wizard. Neither Mark Zuckerberg nor chief technology officer Andrew Bosworth appeared in that top 250.

The board’s 30-day tally exceeded 60 trillion tokens, and the highest-ranked user averaged 281 billion. On the cheapest listed Claude Opus 4.6 rate of $5 for every million tokens, that one user alone could have cost more than $1.4 million. Two days after the board became public, it was gone. A note on the dashboard said it had been meant as a fun way to look at tokens and was shuttered because data had been shared outside the company. Meta said the employee took it down at their own discretion.

Chief people officer Janelle Gale had already told staff that “AI-driven impact” would be a “core expectation” in 2026. OpenAI has run its own employee leaderboard; one power user there burned 210 billion tokens in a week in March 2026. The pattern is the same wherever the score is volume.

THREE COMPANIES, ONE METER

OpenAI’s own explainer puts a token at about four characters per token, with a one-to-two sentence prompt using about 30. That is a tiny unit. Agent loops, always-on helpers, and retries turn it into a large bill because the model is invoked again and again, not because anyone typed a novel.

Uber Burned a Year of Claude Code by April

Uber’s version had no public trophy board, and it still hit the wall. Chief technology officer Praveen Neppalli Naga said the company had used up its entire 2026 Claude Code budget by April, four months into the year. Adoption among engineers classified as agentic coding users jumped from 32 percent in February to 84 percent in March. By spring, 95 percent of Uber engineers used AI tools each month, and about 70 percent of committed code came from those tools.

Average cost sat between $150 and $250 per engineer per month. Power users ran $500 to $2,000. Naga said he spent $1,200 in a two-hour personal demo. In May, chief operating officer Andrew Macdonald said the company could not draw a clean line from that spend to better products for riders and drivers.

That link is not there yet, right? I think maybe implicitly there is more that is getting shipped, but it’s very hard to draw a line between one of those stats and, ‘Okay, now we’re actually producing 25 percent more useful consumer features.’

Andrew Macdonald, Uber chief operating officer, on Rapid Response

Macdonald also said Uber would have to talk about token cost versus headcount if it could not show useful features. That is the adult version of Treadwell’s warning. Use is cheap to count and expensive to fund.

The same pressure shows up outside the coding tools. In March 2026, PwC’s US chief executive Paul Griggs told partners that senior staff who were not “paranoid about being AI-first” would likely be replaced. “I don’t think anyone gets a free pass here. Anyone,” he said. Anyone who thought they could opt out of AI, he added, was “not going to be here that long.” Under that kind of instruction, a token leaderboard is not a toy. It is a shield.

What Kiro Credits Cost on the Public Menu

Kiro is AWS’s agentic coding environment, sold as a spec-first alternative to loose chat-driven coding. On the public site, work is billed in credits, not in a raw token ticker, and the meter still runs. Kiro’s public credit prices start at free and climb with the allowance.

KIRO SELF-SERVE PLANS

Paid plans can buy extra credits at $0.04 each. Enterprise billing, single sign-on, and usage analytics sit behind a sales conversation. Inside Amazon the unit may be a token and not a public credit, but the physics are the same: more agent steps mean more paid work.

That is the twist only Amazon can own. It sells this meter to other companies, and it just told its own staff that climbing the meter is not the job. The product did not get pulled. The scoreboard did.

Treadwell’s Warning Outlived the Scoreboard

By September 2026, AWS vice president of agentic AI Swami Sivasubramanian said more than 100,000 Amazon engineers use Kiro. The unofficial rank list is gone. The tool is not. A Kiro write-up of work on Amazon Bedrock’s inference engine describes six engineers finishing in 76 days a rebuild first scoped as a much larger, much longer job, with about 100,000 lines of new Rust and 95 percent of the code coming through Kiro.

Anthony Liguori, a vice president and distinguished engineer at AWS, said, “To my surprise, what I thought was going to take us over a year with a large team, we were able to ship in just a couple of months with only six people.” That is the outcome Treadwell asked for: a customer problem closed. It is also the kind of story a token rank cannot tell, because a short, sharp project can burn fewer tokens than a week of idle agents.

Kirorank is deprecated. The weekly AI-use target for developers is still the rule those engineers work under. Amazon will keep counting tokens, because tokens are how it knows the bill. It will not, for now, hang that count on a wall with names attached.

Frequently Asked Questions

What Is Amazon Kiro?

Kiro is AWS’s spec-driven coding environment, launched in July 2025, and it now spans a desktop IDE, a command-line tool, and a web agent. It drafts requirements, design notes, and tasks before it writes code, then bills that work in credits on the public plans, with a $20 credit applied when a user first upgrades to a paid tier through social login or AWS Builder ID.

What Is a Token on These AI Leaderboards?

A token is a small chunk of text or data a model reads or writes, and OpenAI’s public guide treats roughly four characters as one token, with a short one-to-two sentence prompt using about 30. Agent runs multiply that count because each retry, file read, and tool call is more tokens, which is why a “usage” rank can jump without a product changing.

Was Kirorank Used in Amazon Performance Reviews?

Amazon said Kirorank was an employee-built beta dashboard, not a formal or approved tool, and that it was never meant to promote AI use for its own sake. The company still tracks tokens to watch cost and efficiency and says it does not use those counts as a developer productivity score.

What Was Meta’s Claudeonomics Board?

Claudeonomics was an unofficial Meta dashboard that listed the top 250 token users among more than 85,000 employees and awarded titles including Token Legend, Cache Wizard, Model Connoisseur, and Session Immortal. Meta said the employee who built it took it down at their discretion after the data was shared externally, and the company did not request that step.

Does Amazon Still Want Developers on Kiro Every Week?

The late-2025 target for more than 80 percent of developers to use AI each week, with Kiro as the recommended tool, was still the stated aim when Kirorank came down, and by September 2026 Amazon’s agentic-AI lead said more than 100,000 Amazon engineers use Kiro. Taking the board offline changed how that use was displayed, not the instruction to use the tool.