TellWorldAI World Archive

Home›News›Release

ReleaseGenerative AI for everyone

Anthropic launches computer use with an upgraded Claude 3.5 Sonnet and Claude 3.5 Haiku

On 22 October 2024 Anthropic announced an upgraded Claude 3.5 Sonnet, a new Claude 3.5 Haiku and computer use in public beta, which lets Claude operate a computer through screenshots, cursor movements and typing.

Hands guide a cursor over a generic desktop interface.

Background

The announcement covered two Claude 3.5 models at once: an upgraded Sonnet and a new Haiku.[1]

What happened

On 22 October 2024 Anthropic announced three things: an upgraded Claude 3.5 Sonnet, a new Claude 3.5 Haiku, and computer use in public beta.[1]

With computer use, developers can direct Claude to use a computer much as a person would. Claude looks at a screen, moves a cursor, clicks buttons and types text.[1]

Anthropic said Claude 3.5 Sonnet was the first frontier AI model to offer computer use in public beta. It also called the capability experimental and, at times, error-prone.[1]

Key details

The upgraded Sonnet scored 49.0% on SWE-bench Verified, up from 33.4%. SWE-bench Verified is a test built from real software problems. On OSWorld, a test of computer tasks, it scored 14.9% in the screenshot-only category and 22.0% when allowed more steps.[1]

Claude 3.5 Haiku scored 40.6% on SWE-bench Verified. It matched Claude 3 Opus on many benchmarks.[1]

Early adopters of computer use included Asana, Canva, Cognition, DoorDash, Replit and The Browser Company.[1]

The US and UK AI Safety Institutes tested the upgraded Sonnet before release.[1]

Computer use was offered through the Anthropic API, Amazon Bedrock and Google Cloud Vertex AI.[1]

Impact

Computer use lets Claude act inside software on a user's behalf, a different kind of task from answering questions or writing text. Anthropic framed the feature as experimental, and named the early adopters listed above.[1]

SWE-bench Verified tests whether a model can fix real problems from software projects. OSWorld tests whether a model can complete tasks in a computer operating system. Screenshot-only means the model received only images of the screen, with no direct access to the underlying software.

Benchmarks like these are a snapshot taken at one time, on one set of tasks. They show what the model did in Anthropic's testing, and they are not a guarantee of results on a user's own computer.

Sources

  1. 1

Cite this page: Anthropic launches computer use with an upgraded Claude 3.5 Sonnet and Claude 3.5 Haiku. TellWorld, https://www.tellworld.net/news/computer-use-2024. Updated 2026-11-08

Related news

View all