Tokenmaxxing: measuring AI by consumption is measuring the bill

// 00

What tokenmaxxing is

the thesis of this edition

Tokenmaxxing is treating AI token consumption as proof of adoption or productivity: usage leaderboards, usage targets, promotion criteria. The name spread through the American business press between April and May 2026, when the internal consumption rankings at Meta and Amazon became public.

In February 2026, Meta's CTO said his best engineer was spending the equivalent of his salary on tokens: "Keep doing it. No limit." In June, Meta announced token budgets per team. Seven months separate "no limit" from the budget (section 04).

When it becomes an adoption target, a ranking or a promotion criterion, consumption starts measuring the bill dressed up as productivity, and people do what anyone does with a target. They hit the target. Whoever measures by the token sees the number rise and reads progress. Whoever pays the bill sees the same number and reads expense.

None of this is an argument for spending less on AI. It is an argument for knowing what the spending bought. At Jump, the yardstick is delivery, not tokens. In client migration projects where we use MigrateMind, Jump's AI migration tool, delivery time drops by as much as 70%. What we track is that time, not how many tokens the tool consumed (section 13).

The numbers in this edition come with source and date, and several of the ones in circulation are wrong. Section 11 lists the ones that fell apart.

// 01

What a token is, and why the bill changed

without this, the rest of the edition makes no sense

A token is the piece of text the AI model reads and writes; a word usually comes to one to three of them. Everything that goes into the model (the question, the file, the history, the code the agent opened) is billed as input, and everything that comes out is billed as output, at a price per million of each.

The analogy I use is the taxi meter. The price per kilometer is the price of the token, and it is falling. What changed was the ride. The chat of 2024 was a short ride: one question, one answer. The agent of 2026 is a driver you send off to solve a problem, who drives around the whole city until it is solved, sometimes with three other cars in parallel. The kilometer got cheaper, and the ride got much longer.

Two ways to pay

Subscription: a fixed fee per person and a usage quota that, when it runs out, locks the tool until the window resets; the limit is on usage. Consumption: every token at list price, with no quota; usage, as Anthropic puts it, only adds to what shows up on the bill, and the limit, if there is one, is on spending.

Corporate plans are moving from the first model to the second. In new Claude Enterprise contracts, the US$ 20 a month seat "doesn't include any usage", and every token is billed separately, since November 2025. GitHub Copilot did the same on 1 June 2026: the Business seat stays at US$ 19, now with US$ 19 in token credits and overage billed separately, because "the current premium request model is no longer sustainable." Individual plans remain subscriptions, and that difference comes back in section 07.

// 02

The price falls, the bill rises

the mechanism

The price per token fell, starting in March. The Ramp AI Index, which tracks the spending of more than 70,000 American companies, showed on 9 September 2026 that the effective price per million tokens fell 41% from the March peak, from US$ 1.15 to US$ 0.68. What the headline leaves out is that on 1 January the price was US$ 0.73. Against January, the drop is about 7%. Epoch AI measured declines of 9 to 900 times a year through 2025, at constant performance.

Consumption per request went up. The OpenRouter study with a16z, covering more than 100 trillion tokens, found that the text sent with each request grew about fourfold through November 2025. At Google, the volume processed per month went from 9.7 trillion tokens in 2024 to more than 3.2 quadrillion in 2026, according to Sundar Pichai at Google I/O on 19 May. The Claude Code documentation says agent teams use "approximately 7x more tokens than standard sessions".

The cost per person went up. The Claude Code costs page said, until early April 2026, that the average cost was US$ 6 per developer per day. Between 8 and 16 April, it changed to US$ 13 per active day and US$ 150 to 250 a month. A spokesperson explained to Business Insider that "there was no pricing or product change": the default model became Opus, the largest in the line, and usage changed.

The point for whoever approves the budget is that the discount on the list price does not reach the bill. The price falls on the cheapest model, while the work that produces results moves to the most expensive one and to longer tasks. On the same Ramp base, from January to August 2026, median AI spending per employee rose 113%, from US$ 5.86 to US$ 12.50 a month. That is the median across all employees, with subscriptions, agents, API and GPU, which is why the figure looks small. The math for a Brazilian company is in section 08.

And the top model did not always get cheaper. At OpenAI, it went from US$ 1.25 per million input tokens in September 2025 to US$ 5, four times as much in nine months. At Anthropic, the new top model costs twice the previous one, and the same text generates "approximately 30% more tokens" on recent models.

Gartner published in June 2026 that, by 2028, "AI coding costs will overtake the average developer's salary", with no methodology. The yardstick, according to heise, is a global average salary of about US$ 2,000 a month, close to that of a mid-level developer in Brazil.

// 03

When consumption becomes the target

the cases

In every case below, the company wanted to speed up AI adoption and chose some form of usage as the yardstick: tokens, log-ins, rankings.

Accenture, February 2026: promotion by log-in. An internal email, confirmed to CNBC, told senior managers that promotion to leadership roles would require "regular adoption" of AI, measured by weekly log-ins to the tools: the shallowest possible way to measure usage.

Nvidia, March 2026: the token as a perk, and then as a floor. At the GTC conference, Jensen Huang said he intends to give engineers, on top of their salary, "probably half of that on top of it as tokens". Three days later, on the All-In podcast: "If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed." People who quote him tend to mix up the perk and the consumption floor.

Amazon, May 2026: the ranking switched off. The target was for more than 80% of developers to use AI every week, and an internal dashboard, Kirorank, scored usage. According to the FT, employees put agents on unnecessary tasks to climb the ranking, and Amazon confirmed that the dashboard "has since been deprecated". According to people who spoke to the FT, the yardstick became "normalised deployments", useful code that reaches production.

Uber, April to June 2026: the budget that ran out in four months. With Claude Code opened up to about 5,000 engineers, the CTO said in April: "the budget I thought I would need is blown away already." In June, a spokesperson confirmed to Bloomberg a cap of US$ 1,500 a month, per agentic coding tool and per employee. The COO said later that he still cannot link that spending to what was delivered.

In Brazil, two opposite designs. In May, Folha (in Portuguese; the quotes are my translation) showed that Magalu allocates AI credits per department, "to avoid putting pressure on employees", with no targets, and measures projects that reach production: more than 250 tested and 180 in production since 2025, according to its AI director, Caio Gomes. The CEO, Frederico Trajano, said he has already invested "a few tens of millions of reais" in tokens, "little compared to our capex". The same story says it saw documents showing that Nubank gives an individual allowance and highlights the heaviest users in an internal newsletter. Nubank declined to comment, and the published figure changed between two versions of the story (section 11), so I do not use the number.

// 04

Meta, from "no limit" to a budget

the case from the inside

Meta went through every phase in less than a year, and each one left a trail in reporting backed by documents or official statements.

Two details apply to any company. The first is that the ranking was not official policy. It was born from one employee, on the intranet, and grew because the environment rewarded whoever was seen using AI. Meta says the "AI Native" labels never entered performance reviews, and a lawsuit by about two dozen laid-off employees claims the opposite. I do not know who is right. What can be said is that, with leadership saying "no limit", a consumption ranking found fertile ground.

The second is what Meta put in its place: spending visibility per team, a formal budget and evaluation by the outcome of the work (quality, speed and complexity). None of the three is sophisticated. All three already existed for every other expense in the company.

// 05

What is lost when the yardstick is consumption

and why it is more than money

The money is the visible part, and it is hard to size from the outside. Estimates for the Claudeonomics month range from US$ 221 million to US$ 900 million, all at list price, with no discount or caching (section 11).

The bigger loss is the signal. A company that measures consumption learns who uses the most, and stops knowing whether the usage was worth anything. DORA, Google's annual research on software delivery, surveyed almost 5,000 professionals in 2025. AI adoption shows up linked to more deliveries and also to more failures in production. The adoption dashboard shows success, the incident dashboard may show something else, and nobody looks at the two together.

And what people feel does not work as a yardstick. In 2025, in a randomised study by METR, experienced developers took 19% longer with AI and, afterwards, believed they had gained 20%. In the 2026 round, the central estimates landed between 4% and 18% faster, with confidence intervals that cross zero, but the gap between perception and measurement held. Consumption and perception rise together, and neither one tells you whether the work got better.

The counterargument

Jensen Huang wants each engineer to consume half of their own salary in tokens, and Frederico Trajano calls Magalu's tens of millions in tokens "little compared to our capex". The argument behind both statements is that spending a lot on AI is cheap compared to the cost of falling behind. I agree with half of it. It holds for the size of the investment, and not for the yardstick. Magalu itself spends that money and measures projects in production.

There is a stronger version. Early on, measuring usage is a legitimate sign that the tool has reached people. I agree, as long as the question is "has it reached them?". When it becomes "was it worth it?", the yardstick has to change, and the moment is when AI spending gets its own budget. From then on, consumption is cost, and what gets measured is what it delivered.

// 06

Why AI spending does not show up in time

the bill arrives as a single number

On 28 July 2026, at an internal meeting reported by the FT, senior Amazon engineers presented three AI budget overruns. The largest was a project that used Claude Sonnet to match author data to product pages: US$ 1.8 million spent, 860% over budget, five months until anyone noticed, and the project did not work. The other two were US$ 541,000 and US$ 134,000.

Five months, at one of the companies that knows the most about cloud costs in the world.

The reason is structural. Servers and disks get tagged with a team and a cost centre; a call to an AI model is a transaction, and the tag only exists if someone puts it on the call. In the words of the FinOps Foundation, the OpenAI or Anthropic invoice "will show spend by API key or project, not by business unit, cost center, application, or team". The most common finding in a first inventory, according to the same document: "engineering teams are using personal or team credit cards to pay provider invoices that no one in Finance has visibility into."

The aggregate numbers show the effect. Ramp's economics blog measured, in April 2026, among companies using its token management product, an average monthly spend of US$ 140,842 and a median of US$ 2,246. The average is sixty times the median. For planning, the average is the wrong number, and it is the one that usually ends up in presentations.

In Brazil, the quarterly KPMG survey with a national cut (second quarter of 2026) says that 58% of companies claim full visibility over technology costs, that 49% already have a specific budget for tokens and that 41% have postponed or scaled back agent initiatives because costs came to exceed the expected benefits. KPMG does not say how many Brazilian companies responded, so the numbers count as a signal, not as a portrait.

// 07

The pattern: the bill that does not go through the company

what links this edition to the agentic systems series

Shadow AI (in Portuguese), the AI that comes into the company without going through whoever approves it, has a cost side, and the pricing shift in this edition reinforces it. The personal plan has a fixed cost and a usage limit; the corporate plan, more and more, has a variable cost and a spending limit. For the same heavy work, the personal subscription goes further, because the vendor subsidises intensive use. SemiAnalysis estimated, in October 2026, that a US$ 200 Anthropic plan still had half its quota left after consuming the equivalent of US$ 2,485 at API prices.

Ramp measured the effect. In February 2026, according to its data published by Axios, 28% of corporate spending on OpenAI's chat went through personal consumer plans, and not through business plans.

That spending disappears twice: from the cost inventory, because it comes in as a card reimbursement, and from the data contract. Under Anthropic's consumer terms, the user chooses whether conversations train models, with five-year retention if they opt in; no training by default and a DPA apply only to Team, Enterprise and the API, and at OpenAI, to the business plans and the API.

In the agentic systems series, the risk came from an authorisation granted outside where the company looks. Here it is the credit card. When the adoption yardstick pushes for usage and the official allowance has a ceiling, usage looks for the path that has none.

// 08

How much does your company spend on AI per person?

the question almost nobody can answer

The honest answer, at most companies, is "I don't know". You have to put together invoices, usage-billed seats, card reimbursements and personal accounts used for work, and the last one is invisible. The public numbers help little:

• Anthropic gives US$ 150 to 250 per developer per month, for Claude Code alone, as a vendor average, with no sample disclosed.
• DX, which sells an engineering measurement platform, estimates US$ 200 to 600 per engineer per month for seat plus tokens, with no sample.
• The open poll of the Gartner community (351 participants, July 2026) shows 22% in the US$ 201 to 500 per developer per month range. It is not a representative survey.
• Ramp measures a median of US$ 46 per employee per month (all employees, not only developers) among customers of its token management product. That is a different group from the index in section 02, which covers the more than 70,000 companies in its base and gives US$ 12.50; the two numbers are not comparable.
• In Brazil, I found no figure for AI spending per company or per person measured with a sample. Cetic.br measures adoption: 17% of companies with ten or more employees used AI in 2025, in a sample of 4,174. Adoption tells you who uses it, not what it costs.

Three scenarios, in reais

With no measured Brazilian number, the math comes from the public ranges. I use the dollar at R$ 5.20, just for the math; redo it with the day's exchange rate, since the order of magnitude does not change. On an international card, add 3.5% of IOF, the Brazilian tax on financial transactions.

• Baseline: 100 developers with Claude Code at Anthropic's average, US$ 150 to 250 per person per month. From R$ 0.94 million to R$ 1.56 million a year.
• Tail: the same 100 at the cap Uber set after the overrun, US$ 1,500 per person per month. About R$ 9.4 million a year.
• Whole company: 2,000 employees at the median of Ramp's customers who manage tokens, US$ 46 per person per month. About R$ 5.7 million a year.

Per developer, US$ 150 to 250 a month is 7.5% to 12.5% of the global average salary Gartner uses as its yardstick; Uber's cap would be 75%. And the cost of a hired developer to the company is well above the salary, so the right comparison, for whoever builds the budget, is with the total cost of the position.

There is no public rule for when the subject goes up to the board. The one I would use has two triggers: annualised AI spending exceeds the amount that, in any other contract, would require committee approval; or it grows for two consecutive quarters without a measured outcome next to it.

// 09

How to limit AI spending, by level of assurance

from what works on its own to what depends on goodwill

The four most widely used tools have a limit that cuts off usage when spending reaches a set amount, and in all of them it has to be switched on.

• Anthropic: a cap per organisation and workspace in the API; in Claude Enterprise, per organisation, group and person. The documentation is literal: "No limit anywhere = no limit."
• OpenAI: "spend alert" only warns; "hard spend limit" cuts off, per organisation and project, with some delay.
• GitHub Copilot: since June, the budget per enterprise, cost centre or user can block AI credits.
• Gemini: a cap per project in the API, with about ten minutes of delay paid for by the customer; in Gemini Enterprise, no per-user limit.

An automatic cutoff works for internal use, testing and developer tools. In an agent serving customers in production, cutting off takes the service down. There, the limit becomes an alert with fast approval and a cheaper fallback model for when the ceiling is reached.

Along with the cap, one API key per team and application, with an owner, a cost centre and an approved use case. It is the FinOps Foundation's first control, and it turns the bill from a single number into spending per team.

• A budget with a measured baseline. The FinOps Foundation recommends measuring for 30 to 60 days before setting any amount, budgeting 110% to 120% of the baseline and alerting at 80% and 100%.
• Cost per outcome. "$42,000 on tokens this month" says nothing; "$0.18 per ticket, down from $0.31" does. Discarded runs are cost, and in agentic systems the infrastructure around the model is often 40% to 60% of the spend, without showing up on the vendor's invoice.
• One outcome metric per use case, defined before scaling; the Stanford study of 51 AI deployments (April 2026) links this to being able to demonstrate value. In engineering, throughput always goes together with failures and rework.

Cost per outcome, outside engineering

Cost per outcome is the total AI spend on a use case, including the surrounding infrastructure and discarded runs, divided by the number of deliveries that succeeded. An example from customer support, with round numbers made up to show the math, not a measurement from Jump or from a client: in 30 days, the model invoice came to US$ 2,400 and the agent closed 20,000 tickets, or US$ 0.12 per ticket by the invoice math. Adding 50% for infrastructure and subtracting the 4,000 tickets that were reopened or handed off to a human agent, it is US$ 3,600 for 16,000 resolved: US$ 0.23 per outcome, almost double.

That is the AI cost per outcome. The business outcome comes from the next comparison: the total cost per resolved ticket, adding AI and the agents' time on the 4,000 handed off, against the cost per ticket before the agent. That number, and its direction month by month, is what goes to the board.

Who is accountable for what

• Finance (CFO or the controller's office): owns the consolidated number and the AI budget, including cards and reimbursements.
• Technology (CIO or CTO): switches on the limits, issues the per-team keys and approves vendors.
• Head of each department: owns the use case and the outcome metric, and is accountable for cost per outcome.
• FinOps, where it exists: consolidates the invoices, allocates by team and triggers the alerts. Exceptions above the cap are approved by whoever holds the department's budget, with Finance informed.

All of this works if someone looks. A dashboard nobody opens has the same effect as no dashboard, and that is how Amazon took five months to see the US$ 1.8 million project.

• Consumption rankings and adoption targets based on usage. Meta and Amazon switched theirs off; DX calls tokens "diagnostic telemetry", which shows how the work changed, not whether it improved.
• Perception surveys ("how much do you think you gained?"). METR showed, twice, that they do not track the measurement.
• Banning personal accounts without a corporate plan that offers the same. It pushes usage out of sight.
• Spending alerts without a cutoff, where a cutoff would be possible. They warn you after the money is gone.

// 10

What to do about AI spending

this week, this quarter, in the contract

This week

1. Take an inventory of every AI vendor account, every API key and every payment method, including personal card reimbursements.
2. At every vendor, switch on the limit that cuts off for internal use and testing. For agents serving customers, an alert with fast approval and a fallback model.
3. In Claude Enterprise, set limits per group and per person. Without configuration, there is no limit.
4. Take the per-person token consumption ranking off every dashboard.
5. Define who is accountable for AI spending, with the authority to enforce the rule (who is accountable for what).

This quarter

1. One key per team and per application, with an owner, a cost centre and an approved use case.
2. Measure for 30 to 60 days before setting a budget. Then budget on the measured baseline.
3. For each use case, choose an outcome metric before scaling, and track throughput together with failures and rework.
4. Report cost per successful outcome, counting discarded runs and the infrastructure around the model.
5. Review the model used for the most expensive workloads. The smallest model that passes the quality bar is the optimisation with the biggest effect, according to the FinOps Foundation.
6. Set the amount above which AI spending moves off the team's card and goes through procurement.

In the contract

1. Work with company data only on a corporate plan, with a DPA and no training by default. If the corporate plan has a cap and the personal one does not, review the cap before banning the personal one.
2. Before signing with a vendor abroad, ask your tax counsel to classify the service: Brazil's Federal Revenue Service has no settled position on AI APIs, and the tax changes depending on whether the contract is read as a license, SaaS or a technical service. On an international card, IOF is 3.5%.
3. Volume commitments only with 60 to 90 days of measured usage. Reserved capacity can come out more expensive per token than pay-as-you-go.

A question for anyone who made it this far: at your company, who is accountable for AI spending today? Finance, technology, each department, or nobody?

// 11

What circulates and does not hold up

each item with its source

// 12

What was not verified

what I looked for and could not confirm

• Meta's memos are not public; what is here comes from reporting by outlets that saw the documents and from spokespeople.
• Dave Treadwell's line at Amazon ("Please don't use AI just for the sake of using AI") was reported by people who were present, with no recording. The "normalised deployments" metric comes from anonymous sources and was not confirmed by Amazon.
• The real cost of Claudeonomics and Meta's internal AI cost in 2026 are not public beyond "in the billions".
• The financial reason for Microsoft's Claude Code cut comes from anonymous sources.
• The exact minute of Jensen Huang's remark on All-In was not checked against the video.
• The number of Brazilian companies in the KPMG survey is not disclosed, and, as far as I could find, there is no figure for AI spending per person in Brazil measured with a sample.
• The subsidy on personal plans by model only appears in a SemiAnalysis chart that I did not see in full, and the Axios story attributed to Aaron Levie could not be located.
• The taxation of AI APIs contracted abroad has no settled position from Brazil's Federal Revenue Service. This edition does not publish a combined tax rate.

// 13

Jump's picture

from here on I talk about what we do at Jump

Two questions run through this edition: how much the company spends on AI, and where; and who uses it, with which tool and at what risk. At Jump, each one has a product.

What exists today

how much and where Opt3, Jump's FinOps platform. It gives visibility and governance over cloud costs on AWS, Azure, GCP and Databricks, separates AI cost (Bedrock, Azure OpenAI, Vertex, Databricks) from the rest of the bill and also reads direct invoices from OpenAI and Anthropic.

who and with what RADAR·AI. It discovers AI usage, including undeclared usage, through connectors that read only metadata, and puts each system in an inventory with an owner and a risk level. It does not read spending. Opt3 shows the bill, and RADAR·AI shows who is behind it.

How we measure: delivery, not tokens

MigrateMind is Jump's AI migration tool. It migrates data and analytics workloads off legacy platforms, with real execution and validation between source and target. In client projects where we use MigrateMind, delivery time drops by as much as 70%. The gain varies from project to project, and 70% is the ceiling we have seen, not the average. The yardstick is delivery, measured as the time until the client receives the validated migration, and not the tokens the tool consumed along the way.

What is on the roadmap, stated as roadmap

The limit, said out loud

// 14

Primary sources

primary source whenever accessible; when read through a republication or citation, the entry says so

• Forbes · The ‘AI Gods’ Spending As Much As They Can On AI Tokens · 31.03.2026
  forbes.com
  Andrew Bosworth's remarks at the February event in San Francisco: the best engineer spends the equivalent of their salary on tokens, and "No limit".
• The Information · Meta Employees Vie for AI ‘Token Legend’ Status · 06.04.2026; with Fortune, A Meta employee created a dashboard so coworkers can compete to be the company's No. 1 AI token user, 09.04.2026, and The Pragmatic Engineer, The Pulse: ‘Tokenmaxxing’ as a weird new trend, 23.04.2026
  theinformation.com · fortune.com · blog.pragmaticengineer.com
  Claudeonomics: a dashboard built by one employee, more than 85,000 people, 60.2 trillion tokens in 30 days and 281 billion for the person in first place. Fortune records the message that took the dashboard down. The original report is behind a paywall and was read through the quotations.
• The Decoder, citing The Information · Meta shifts from "tokenmaxxing" to token managing as internal AI costs reportedly hit billions, 13.06.2026; Meta drops AI usage from engineer performance reviews after "tokenmaxxing" backfires, 08.09.2026
  the-decoder.com · the-decoder.com
  The June memo to about 6,000 employees: costs "in the billions", the AI Gateway and token budgets from 2027. Bosworth's line about tokens and impact. The "quality, speed and complexity" criteria in the September memo.
• WIRED · Meta Pushes Its New AI Agent on Employees—but Eases Off on Tokenmaxxing · 02.09.2026
  wired.com
  The new review guidance without "AI usage", the warning that adoption dashboards and token counts do not measure impact, the spokesperson's statement on the "AI Native" labels, the November 2025 announcement and the lawsuit by the employees dismissed in May.
• Ramp · September 2026 Ramp AI Index: Cracks in the AI thesis, part 2 · 09.09.2026
  ramp.com · datawrapper.dwcdn.net (data behind the price chart) · datawrapper.dwcdn.net (data behind the spend-per-employee chart)
  The effective price per million tokens: $0.73 in January, $1.15 at the March peak, $0.68 in September. Median AI spend per employee rising from $5.86 to $12.50 between January and August, with subscriptions, agents and GPUs included. The figures come from the data files Ramp itself used in the charts.
• Anthropic · Manage costs effectively (Claude Code documentation) · read on 13.10.2026
  code.claude.com · web.archive.org (copy of 08.04.2026, with $6) · web.archive.org (copy of 16.04.2026, with $13)
  The average cost estimate that went from $6 per day to $13 per active day, the range of $150 to 250 per month and the "approximately 7x more tokens" for agent teams. The two archived copies date the change to between 8 and 16 April 2026.
• Anthropic · How am I billed for my Enterprise plan? (help centre); Plans & pricing · read on 13.10.2026
  support.claude.com · claude.com
  The $20-a-month Claude Enterprise seat that "doesn't include any usage", every token billed at API prices, the end of seats with bundled usage in new contracts and usage that "only increases what shows up on the bill". The pricing page confirms that Pro, Max and Team remain monthly subscriptions.
• Gizmodo · Anthropic Hiked the Price for Power Users Amid Complaints Its Model Is Getting Worse · 15.04.2026, corrected 17.04.2026
  gizmodo.com
  The Anthropic spokesperson says usage-based billing in Enterprise began in November 2025. The correction of 17.04 undoes the version that treated the change as new.
• GitHub · GitHub Copilot is moving to usage-based billing (changelog and post by Mario Rodriguez) · 27.04.2026
  github.blog/changelog · github.blog
  The $19 Business seat with $19 in credits spent per token from 1 June, overage at the published price and the two sentences of the rationale.
• Epoch AI · LLM inference prices have fallen rapidly but unequally across tasks · 12.03.2025
  epoch.ai
  The fall in price at constant performance, from 9 to 900 times a year depending on the capability level, with data up to early 2025.
• OpenRouter and a16z · State of AI: An Empirical 100 Trillion Token Study · 04.12.2025
  openrouter.ai
  More than 100 trillion tokens over 13 months: text per request from about 1,500 to more than 6,000 tokens, and programming from about 11% to more than half of all tokens. The data comes from one platform, not the whole market.
• Google · I/O 2026: Welcome to the agentic Gemini era, Sundar Pichai · 19.05.2026
  blog.google
  Tokens processed per month across all Google products: 9.7 trillion two years ago, 480 trillion at I/O 2025 and more than 3.2 quadrillion in 2026.
• Business Insider, republished by Yahoo Finance · Anthropic doubles its estimate of what Claude Code tokens will cost engineers · 29.04.2026
  finance.yahoo.com
  The Anthropic spokesperson's statement, "there was no pricing or product change", and the explanation via the switch of the default model to Opus 4.7.
• OpenAI · Pricing (API pricing page) · copies of 01.04.2023, 18.09.2025 and 01.06.2026
  web.archive.org · web.archive.org · web.archive.org (openai.com refuses automated reading; read from archived copies)
  The input price of the top model: $30 per million for GPT-4, $1.25 for GPT-5 and $5 for GPT-5.5.
• Anthropic · Pricing (API documentation), read on 13.10.2026; Claude Fable 5 and Claude Mythos 5, 09.06.2026
  platform.claude.com · anthropic.com
  Opus stepping down from $15 to $5 and $4 per million input tokens, Fable at $10, and the warning that, from Claude 4.7 onwards, the same text produces "approximately 30% more tokens".
• Gartner · Gartner Predicts AI Coding Costs Will Surpass Average Developer’s Salary by 2028 as Token Consumption Surges · 24.06.2026
  gartner.com
  The 2028 forecast and the sentence quoted in the text. The press release gives neither sample nor methodology.
• heise online · Prognose: 2028 wird KI-Coding teurer als Entwicklergehälter · 25.06.2026
  heise.de
  The yardstick behind the forecast: a global average salary of about $2,000 a month.
• Financial Times and CNBC · Accenture combats AI refuseniks by linking promotions to log-ins; Accenture tells senior staff to use AI tools or risk losing out on leadership promotions · 19.02.2026
  ft.com · cnbc.com
  The email linking promotion to "regular adoption" of AI, the collection of weekly log-ins, the spokesperson's confirmation to CNBC and the exemption for 12 European countries.
• CNBC · Nvidia's Huang pitches AI tokens on top of salary as agents reshape how humans work · 20.03.2026
  cnbc.com
  Jensen Huang's remarks at GTC: tokens worth "probably half" of salary, on top of it, as a benefit.
• Business Insider · Jensen Huang says he would be 'deeply alarmed' if his $500,000 engineer did not consume at least $250,000 of tokens · 20.03.2026; All-In Podcast episode of 19.03.2026
  businessinsider.com · youtube.com
  Huang's second remark, three days later, in which tokens become a consumption floor for a $500,000 engineer.
• Financial Times · Amazon scraps AI leaderboard to stop workers chasing usage scores · 28.05.2026; with InfoWorld, Amazon deletes devs’ tokenmaxxing leaderboard to minimize costs, 29.05.2026
  ft.com · ftchinese.com (full text in English on the FT Chinese mirror) · infoworld.com
  The target of more than 80% of developers using AI every week, Kirorank, Amazon's confirmation that the leaderboard "has since been deprecated", the sentence attributed to Dave Treadwell and the "normalised deployments", the latter from anonymous sources.
• The Information · Uber CTO Shows How Claude Code Can Blow Up AI Budgets · 04.2026; with TEXXR, 15.04.2026
  theinformation.com · texxr.com
  The rollout of Claude Code to about 5,000 engineers, usage going from 32% to 84% and Praveen Neppalli Naga's sentence about the blown budget. The report is behind a paywall; the sentence was checked against TEXXR.
• Bloomberg · Uber Caps Usage of AI Tools Like Claude Code to Manage Costs · 02.06.2026
  bloomberg.com · simonwillison.net (excerpt reproduced on 03.06.2026)
  The Uber spokesperson's confirmation: a cap of $1,500 a month, per agentic coding tool and per employee.
• Rapid Response (Masters of Scale) · Uber's swerve on gas prices, hotels & a driverless future · 22.05.2026
  mastersofscale.com
  The interview with Andrew Macdonald, Uber's chief operating officer, with the transcript: "that link is not there yet" between token spend and useful features shipped.
• Folha de S.Paulo · Empresas distribuem vale IA a funcionários para teste de ferramentas inovadoras [Companies hand out AI vouchers to employees to test innovative tools] · 17.05.2026, amended 18.05.2026
  folha.uol.com.br · web.archive.org (original version, archived copy of 17.05.2026)
  Magalu's credits by department, with no targets, and the statements by Frederico Trajano and Caio Gomes. Nubank with an individual allowance and an internal bulletin. The archived copy shows the "monthly" that the current version changed to "per week", with no correction note.
• The Wall Street Journal · Why Some Companies Say AI 'Tokenmaxxing' Is Key to Survival · 14.04.2026
  wsj.com
  The Meta spokesperson: the dashboard came down at the discretion of the employee who built it, and the company has another dashboard, "AI Insights". Read in part; the text is behind a paywall.
• Exame · Meta transforma uso de IA em disputa interna por produtividade entre funcionários [Meta turns AI use into an internal productivity contest among employees] · 11.04.2026
  exame.com
  The origin of the $900 million, which the report itself calls an estimate based on Claude's public pricing.
• Google DORA · Announcing the 2025 DORA Report; State of AI-assisted Software Development 2025 · 23.09.2025
  cloud.google.com · dora.dev
  Nearly 5,000 professionals surveyed. AI adoption has a positive relationship with throughput and a negative one with stability. 90% use AI, and more than 80% say they are more productive.
• METR · Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, 10.07.2025; We are Changing our Developer Productivity Experiment Design, 24.02.2026
  metr.org · metr.org
  The 2025 randomised study: 16 developers, 246 tasks, 19% more time with AI and a 20% perceived gain. The 2026 round: 57 developers, an effect of between 4% and 18% faster with a margin that includes zero, and the authors' caveat about the study design.
• Financial Times, reproduced by Yahoo Finance and The Next Web · Amazon blew $1.8 million on a failed AI task — and didn't notice for 5 months; Amazon spent $1.8m on a Claude job that failed, and it sells the fix · 30.07.2026
  finance.yahoo.com · thenextweb.com (the FT text is behind a paywall)
  The 28 July meeting: $1.8 million on a project with Claude Sonnet, 860% over budget and five months before anyone noticed. The overruns of $541,000 and $134,000.
• FinOps Foundation · Tokenomics: Managing AI Value in SaaS Model Token Costs (working group) · read on 13.10.2026, with prices current as of May 2026
  finops.org
  The invoice by key or project, with no cost centre. Limits in tokens per minute. The credit cards finance does not see. The order of controls: inventory, a key per team, a budget on a 30- to 60-day baseline, cost per outcome, 40% to 60% of infrastructure around the model and model tuning as the biggest saving.
• Ramp · How much do AI tokens cost businesses?, Rafael Hajjar · 08.06.2026
  ramp.com
  In April 2026, among customers of Ramp's token management product: average monthly spend of $140,842, median of $2,246 and $46 per employee. The number of companies is not given.
• KPMG Brasil · Global AI Pulse Q2 2026 · 28.09.2026
  kpmg.com
  The Brazilian cut: 58% with full visibility of technology costs, 49% with a dedicated token budget and 41% that postponed or scaled back agent initiatives because of cost. KPMG does not say how many Brazilian companies responded.
• SemiAnalysis · Anthropic Subscriptions Offer 5x+ More Value Than OpenAI · 05.10.2026
  newsletter.semianalysis.com
  Anthropic's $200 plan with half its quota left after consuming the equivalent of $2,485 at API prices. Part of the text is behind a paywall.
• Axios · Don't get used to cheap AI · 12.03.2026
  axios.com
  The Ramp figure: in February, 28% of corporate spend on OpenAI's chat went through personal consumer plans.
• Anthropic · Updates to Consumer Terms and Privacy Policy · 28.08.2025
  anthropic.com
  On the Free, Pro and Max plans, including Claude Code used under those accounts, the user chooses whether conversations train models, with five-year retention if they accept. The rule does not apply to Team, Enterprise or the API.
• OpenAI · Enterprise privacy at OpenAI · updated 08.01.2026
  openai.com
  The plans covered by the enterprise commitments: Business, Enterprise, Edu, Healthcare and API.
• DX · AI coding assistant pricing and ROI guide (2026): costs, benchmarks, and what the data shows, Taylor Bruneaux · 12.06.2026
  getdx.com
  The estimate of $200 to 600 per engineer per month, adding seat and tokens. It is an estimate from a company that sells a measurement platform, with no sample.
• Gartner Peer Community · We asked this question in March 2026 and are revisiting it to track changes in AI token spend over time. How much do you spend monthly per developer on tokens for AI-powered tools (e.g., Claude Code)? · 17.07.2026
  gartner.com
  The open poll with 351 participants: 22% between $201 and 500 per developer per month and 2% above $2,000. Participants are self-selected.
• Cetic.br · TIC Empresas 2025 [ICT Enterprises 2025], executive summary and indicator H9 · 2026
  cetic.br · cetic.br (table)
  17% of Brazilian companies with ten or more employees used AI in 2025, against 13% in 2024, in a sample of 4,174. The survey measures adoption, not spend.
• Anthropic · Rate limits (API documentation); Claude Enterprise consumption guide (help centre) · read on 13.10.2026
  platform.claude.com · support.claude.com
  The spend cap per organisation and per workspace in the API, with calls refused once it is reached. In Enterprise, the limit per organisation, group and person, and the rule "No limit anywhere = no limit".
• OpenAI · Spend limits (API documentation) · read on 13.10.2026
  developers.openai.com
  The "spend alert", which only warns, and the "hard spend limit", which cuts off, per organisation and per project. The caveat that the cut-off is not instantaneous.
• GitHub · Budgets and alerts (GitHub Docs) · read on 13.10.2026
  docs.github.com
  The budget per enterprise, cost centre, organisation or user that blocks AI credits. For what is billed per licence, the budget only warns.
• Google · Giving you more transparency and control over your Gemini API costs, 16.03.2026; Overview of overages and spend controls (Gemini Enterprise), read on 13.10.2026
  blog.google · docs.cloud.google.com
  The per-project cap in the Gemini API, with a delay of about ten minutes and any overage charged to the customer. The absence of a per-user limit in Gemini Enterprise.
• Stanford Digital Economy Lab · The Enterprise AI Playbook, Elisa Pereira, Alvin Wang Graylin and Erik Brynjolfsson · 04.2026 (draft)
  digitaleconomy.stanford.edu
  The study of 51 successful deployments, which links a KPI defined before deployment to being able to demonstrate value. The PDF carries a draft watermark.
• DX · Revisiting the DX Core 4 in the Age of AI, Brian Houck · 24.06.2026
  newsletter.getdx.com
  Adoption, token usage and tasks handed to agents as "diagnostic telemetry": they show how the work changed, without saying whether it improved.
• Presidência da República · Decreto nº 6.306, de 14 de dezembro de 2007 [Decree No. 6,306 of 14 December 2007], art. 15-B, VII, as worded by Decree No. 12,499/2025 · read on 13.10.2026
  planalto.gov.br
  The 3.5% IOF tax on the foreign exchange that pays for card purchases of goods and services abroad.
• CISO Advisor · 'Tokenmaxxing' divide empresas sobre métrica de uso de IA entre funcionários ['Tokenmaxxing' divides companies over employee AI usage metrics] · date uncertain (the page prints 04/10/2026 and cites the WSJ of 14.04.2026)
  cisoadvisor.com.br
  The origin of the "$1.8 million" attributed to a Meta employee: the text says the volume of the person in first place "could represent" that bill.
• The Verge · Microsoft starts canceling Claude Code licenses, Tom Warren · 14.05.2026
  theverge.com
  Rajesh Jha's memo withdrawing most Claude Code licences in the Experiences + Devices division by 30 June. The text gives no figure per engineer, and the financial motive comes from anonymous sources.