Pictures from the home page
How we use AI, in one glance.
How tokens get burned
The first table is this site's estimate of a person reading what the model wrote. The second table is money and vendor token counts. They are not the same pile.
| Pace | Text passed | Details checked | Tokens burned | Per checked token |
|---|---|---|---|---|
| Thorough, 200 words a minute | 16,000 | 16,000 | 240,000 | 15 |
| Skim, 450. Carver's skim gear. | 36,000 | 16,000 | 540,000 | 34 |
| Scan, 600. Carver's scan gear. | 48,000 | 16,000 | 720,000 | 45 |
| Speed reading, 1,000. This site's row. | 80,000 | 16,000 | 1,200,000 | 75 |
Method: tokens of text equal 60 times words a minute times 1.33, rounded to the nearest thousand. Checked details stay at the thorough count once the reader is faster. Burned tokens are text tokens times 15. The 15 is a round figure, not a GDPval measurement. Duggan and Payne (2009): skimming caught important ideas at 83%, against 67% for reading half the text normally in the same time. Details showed no reliable gain. Picture: four paces.
| Figure | What was counted | Limit |
|---|---|---|
| 2.1% | All US work hours workers said AI saved, August 2026. May of 2026 was 2.2%. | Workers' own estimates. St. Louis Fed via FRED. |
| 5.3% | All US work hours AI assisted, August 2026. May of 2026 was 6.3%. November 2024 was 4.1%. | A different series from hours saved. |
| $201.9 billion | Property and equipment, Microsoft, Alphabet, Amazon, and Meta, July to September 2026. Up 108% from $97.3 billion. | Includes offices and warehouses. Amazon 120%, Alphabet 114%, Microsoft 103%, Meta 80%. |
| $2.67 trillion | Gartner's forecast of worldwide AI spending in 2026, up from $1.79 trillion. Infrastructure $1.48 trillion, services $576 billion, software $462 billion. | A forecast. Not the capex quarter above. Estimate |
| Over 3.2 quadrillion | Tokens a month Google said it processes across its surfaces, at I/O in May of 2026, up from about 480 trillion a year earlier. | One company's count. Not comparable with anyone else's tokens. |
| $0.64 | Blended price per million OpenAI and Anthropic tokens on Ramp, week ending October 11, 2026. Down 25.4% from $0.86. Token spending rose 23.0%. | Ramp's customers only. Anthropic was 60.1% of that spending, OpenAI 36.4%. |
| 10% to 20% | This site's estimate of oversight as a share of companies' agent spending by 2030. | The 10% is Gartner's floor for guardian agents. The 20% is the top of this site's projection. Estimate |
By use case
Writing and lookup lead the logs and the surveys. The surveys do not share a base, so the shares are not a ranking.
| Figure | Use | Base |
|---|---|---|
| About 40% | Writing, among work-related messages on consumer ChatGPT. About two thirds of writing requests, work and non-work, rework the user's text. | Chatterji et al., NBER w34255. |
| Almost a quarter | Editing text, among Microsoft 365 Copilot Chat prompts. Lookup and refining were almost 60% that week. | One week of February 2026. Counts et al. |
| 35% | Computer and math tasks, among Claude.ai conversations. | February 5 to 12, 2026. Anthropic's logs. |
| 37%, 32%, 32% | Search, writing, and generating ideas. Ever used at work. Do not add. | All US workers, Census, March of 2026. No time window. |
| 51%, 49%, 39% | Writing and editing, search or research, general help. | US employees who already use AI at work. Gallup, May of 2026. |
| 22% vs 8% | Coding assistance, frequent users against less frequent users. Automation was 21% vs 8%. Task management was 21% vs 9%. | Employees who use AI at work. Gallup, May of 2026. |
| 73% | Non-work messages on consumer ChatGPT, up from 53%. | June 2024 to June 2025. Chatterji et al. |
| 82% to 89% | Microsoft's estimate of enterprise Copilot Chat use that is work, against about 27% work use on consumer ChatGPT in OpenAI's study. | Different products. A 2026 paper for the Copilot range. |
| Fewer than 3% | Job tasks above 50% adoption among workers who do them. None above 70%. | St. Louis Fed. How far into the job, not which task fills the chat. |
| 37% faster | A short writing task, early 2023, 444 professionals given ChatGPT. Graders scored the work higher too. | One experiment. Experienced open-source developers later took 19% longer with AI tools. METR's 2026 update calls the newer speedup evidence weak. |
| 2% | US workers who said all or most of their work is done with AI. 21% said at least some. September 2025. Unchanged at 2% from October 2024. | Pew. Not a share of chat messages. |
By channel
Where people meet the tool, and what the seller counts. A seat, a registered agent, and a share of adults are three units.
| Figure | Channel | Base |
|---|---|---|
| 60% and 49% | Read AI summaries in search. Use a chatbot. | US adults, February 2026, Pew. Not mutually exclusive. |
| Over 2.5 billion | Monthly users of AI Overviews, Google's figure. | Vendor count. Not a Pew share. |
| 61% | AI users at work who used a chatbot or virtual assistant. | Gallup, third quarter of 2025. Does not split built-in from standalone. |
| 61% | Copilot users who open it inside Word, Excel, or Teams. | Ipsos, March of 2026. About one adult in nine. Mixes paid Copilot and the free app. |
| Over 30 million | Paid Microsoft 365 Copilot seats. | July 2026. A seat is not use. |
| Nearly 40 million | Agents registered in Agent 365. | Registered is not an agent that worked. |
| 62.1% | Engaged GitHub Copilot Business and Enterprise users still mainly on code completions. | July 22, 2026. Not a trial. |
| 6× and 7× | Heaviest Enterprise worker's messages, and frontier firms' messages to Custom GPTs. | Already customers. OpenAI, December 2025. About 2× messages per seat. |
| 76% | Fin's resolution rate, and the same numeral in Salesforce's release. | Conversations Fin takes part in, versus all support volume. Not two measurements. |
| 55% and 52% | CISOs reporting agent-to-application integrations in production, and those comfortable with a person approving security agents. | 100 CISOs, companies over 1,000 employees. a16z. |
On the home page, outside those three
Breadth, frequency, industry, firms, and memory exams that are not the LongMemEval pair. Each row keeps its own question.
| Figure | Reading |
|---|---|
| 52% | US employees used AI at work at least a few times a year, May of 2026, up from 21% in May of 2023. Gallup. Census, no time limit, March of 2026: 56% of US workers. New York Fed, November 2025: 39% of employed adults used AI now or in the past year. |
| 15% | US employees used AI at work daily, May of 2026, Gallup. It was 4% in May of 2023 and still 4% a year later, then 8% in May of 2025. Among employees who use AI, daily users rose from 19% to 29% by this site's count. |
| 11.5% | Employed US adults using generative AI every workday, third quarter of 2026. The St. Louis Fed series has stayed between 9% and 14% since 2024. Gallup counts employees and any AI. The Fed counts employed adults and generative AI. |
| 42% | US technology employees using AI daily, May of 2026, up from 23% a year earlier. Seven of the other nine industries Gallup reports were at 9% to 15%. |
| 7.2% | US job postings on Indeed that mentioned AI, September 2026, double the 3.6% of September 2025. |
| 19% | Employment of 22- to 25-year-olds in the most AI-exposed jobs, below trend on one Stanford study's August 2026 measure. This site has not checked that study at its source. The evidence is contested. |
| 24.6% | US businesses that used AI in the two weeks to October 18, 2026, up from 17.3% in November 2025. Census. Ramp: 55.9% of its customers paid for AI in September 2026. Paying and using are different bases. |
| 47% and 30% | US employees whose organization had integrated AI, and frequent AI users, May of 2026. Gallup. |
| 37% and 39% | Respondents who said AI contributed to EBIT, May and June 2026, against 2025. McKinsey, which sells AI consulting. In mid-2026 surveys with central banks in seven countries, 83% of executives reported no productivity effect over three years. |
| 40% and 22% | Organizations with $1 billion or more in revenue where respondents said they were scaling AI agents, up from 27% in 2025. Smaller organizations were flat at 22%. McKinsey, May and June 2026. |
| 57.7% and 91.8% | LongMemEval. The other memory exams, including the 61% leak, are on the memory picture's table. They are different tests. |
How this pass was judged
The first drawings were small charts under a headline that had already said the thing. This pass was judged on the first screen: is the picture the page, can you see the comparison before you read a paragraph, and does anything collide.
- One token checked, then the burn. Design 9. Fifteen white squares against thirty-four red ones, same unit, bottoms aligned. The number sits above the pile.
- The job stops short. Design 8. The red band is the thin region above 50%, and the words "No task" fill the region past 70%.
- A reread beat the memory. Design 8. Two bars, one scale, 57.7% and 91.8% of the same 97 questions.
- Search, then three blank measures. Design 8. 60% and 49% are the size of the band. The columns under them are empty because those measures are not published.
- 76%. Design 8. One numeral on a black field. The red line is the wider base in the Salesforce release.
- Four paces. Design 8. The pile grows from 15 squares to 75. The last pile is the red one.
Usefulness still requires the denominator on the picture, and the tables under each one hold the readings that do not share that scale. The home page these numbers come from is version 2026.11.02, November 2, 2026. About our data.