AI sets a new record every two months. Three in four U.S. firms still don’t use it.
Four public data sets show how fast AI is moving in the lab, in annual reports, inside businesses and in the economy. The four speeds are far apart.
Before you read further, try something. Draw what you think happened to the share of American businesses using AI over the past year.
That’s one speed. Here are the other three, and then the first one again with a closer look.
First, a unit of measure
What is the Epoch Capabilities Index? Think of it as an IQ test for AI models. Epoch AI, an independent research group, takes scores from dozens of benchmarks and combines them into one number. The benchmarks range from grade-school math to software engineering tasks and graduate-level science questions. Harder benchmarks count for more.
The scale is pinned to two models: Claude 3.5 Sonnet sits at 130 and GPT-5 at 150. A 10-point gain means the same amount of progress anywhere on the scale. When Epoch launched the index, a 5-point gain roughly matched a doubling in the length of tasks an AI agent could finish on its own.
A score alone is abstract, so it helps to translate it back into test results. Drag the slider to see what a model at each level would be expected to score.
Start with what the models can do. Epoch AI scores major models on one scale, the Epoch Capabilities Index. GPT-4 scored 125.9 when it came out in March 2023.
Since then, the top score has been broken 22 times. That works out to a new record about every two months.
The biggest single jump came with OpenAI’s o1 in December 2024, which added 6.1 points. The current record, 167.3, belongs to Anthropic’s Claude Opus 5.5, released in September.
Free models trail by months. The best open-weight model released between July and September scores 157.5, higher than any model released before February 2026.
Public companies noticed. In 2022, two 10-K annual reports used the phrase “generative AI,” according to SEC filings. This year, 1,193 did.
“AI agent” is on the same path. It appeared in one annual report in 2024 and in 58 this year, close to where “generative AI” stood in 2023.
Use is a different story. Every two weeks the Census Bureau asks businesses whether they used AI to produce goods or services. In November 2025, 17.3 percent said yes. By September, 23.8 percent did.
The largest firms are pulling away. Among firms with 250 or more employees, use rose from 30.5 percent to 44.0 percent in the Census data. The gap between them and the smallest firms grew from 13.5 points to 20.2.
In the economy, the signal is faint. Labor productivity grew at an average annual rate of 2.5 percent from 2023 through mid 2026, according to the Bureau of Labor Statistics, compared with 1.9 percent from 2017 to 2019.
Employment in professional and technical services grew 7 percent from 2017 to 2019, by the bureau’s count. It has been flat since mid 2023.
These numbers don’t prove that AI caused the change in productivity or jobs. They measure how far apart the four speeds are.
Which industries are using it
Nearly every industry moved. Information firms lead at about 45 percent, followed by professional services and finance. Manufacturing gained the most, about 8 points. Hotels and restaurants, mining, and transportation sit near 10 percent.
Three more ways to see the gap
How people use Claude
If yours is one of the three in four firms that doesn’t use AI yet, the capability is already there. The best open-weight model scores higher than anything released before February. What’s missing is the work of putting it into one process and measuring what changes.