Skip to content
The Shape of Intelligence

Claude learns to use a computer

Anthropic releases a model that looks at a screen, moves a cursor and types; the assistant becomes an agent, and the year of agents that follows starts here.

category
product
significance
3 of 5
people
Dario Amodei
organisations
Anthropic

what had to happen · 54 events back to 1943

Every event this one built on, transitively, in order. Direct influences are marked.

VI · Everyone · 3

  1. 2022InstructGPT
  2. 2023Claude
  3. 2024Claude 3 catches GPT-4direct

On 22 October 2024 Anthropic released an upgraded Claude 3.5 Sonnet with a capability it called computer use, in public beta. Given screenshots, the model could work out where things were on the screen, move a cursor, click, scroll and type, and so operate any software written for people: fill in a form from a spreadsheet, look something up in a browser, run a command. It was slow, made mistakes a person would not, and on the company's own benchmark completed about fifteen percent of the office tasks it was set. It was also the first general-purpose agent from a frontier laboratory that acted rather than advised.

The release framed the next eighteen months. OpenAI's Operator followed in January 2025, Google's Project Mariner and the browser agents of 2025 after it, and the coding agents, Claude Code and Codex, that by 2026 wrote a large share of the world's new software began as this idea applied to a terminal.

Computer use is on this timeline as the point at which the models got hands. The organism on this page, in the final chapters, is wired to tools for the same reason.

what it led to · 5 events downstream, through 2026

Built on it directly:

  1. 2024The Model Context ProtocolVI
  2. 2025Claude 4 and Claude CodeVII

And, through them, by era:

sources · 1

See this era in the exhibition →Back to the timeline