· Oct 22, 2024
Anthropic describes Claude’s computer-use capability
Markdown · JSON · Original source
’s article, published October 22, 2024, describes using a computer through screenshots, movements, clicks and keyboard input. Anthropic argues that this could let AI systems use ordinary software rather than relying on tools specially designed for them. The article’s central idea is to make the model fit the computer environments people already use; Anthropic presents computer use as bringing Claude’s existing abilities to computer tasks, not fundamentally increasing those abilities. Anthropic says Claude learned to interpret screens and act on them, and reports that it generalized from training on simple software. On OSWorld, Anthropic reports a result below human performance but above the next-best AI model in its category. The evaluation does not establish how reliably Claude handles everyday work. The article warns that malicious instructions in screen content could divert Claude from a user’s intent, and that users could misuse computer actions. Anthropic says computer use remains slow and error-prone, with some ordinary actions unsupported. It expects the capability to improve and says researchers will work with safety teams at every stage. How well safeguards prevent unintended actions remains an open question.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
· Article
Source details
Original source date: Oct 22, 2024 (2024-10-22T00:00:00.000Z).
Date basis: Date stated on the publisher’s page.
Added to Escape Meridian: 2026-11-16T18:48:25.243Z. First observed: 2026-10-22T08:23:14.075Z. Retrieved: 2026-10-29T06:31:36.490Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on October 29, 2026; originally published October 22, 2024. ’s OSWorld result is an evaluation score; the article does not establish how reliably performs everyday work.
Open question
How well will safeguards prevent screen content or user requests from leading to take unintended actions?
Source-stated
Anthropic says Claude can operate a computer through screen-based actions when given the appropriate software setup.
follow a user’s commands to move a cursor around their computer’s screen, click on relevant locations, and input information via a virtual keyboard
Source-stated
Anthropic argues that computer use could let AI systems work with ordinary software rather than relying on bespoke tools.
The next frontier is computer use: AI models that don’t have to interact via bespoke tools, but that instead are empowered to use essentially any piece of software as instructed.
Source-stated
Anthropic reports Claude scored below human performance but above the next-best AI model on OSWorld.
On one evaluation created to test developers’ attempts to have models use computers, OSWorld, Claude currently gets 14.9%. That’s nowhere near human-level skill (which is generally 70-75%), but it’s far higher than the 7.7% obtained by the next-best AI model in the same category.
Source-stated
Anthropic identifies prompt injection as a risk when Claude interprets screenshots from internet-connected computers.
Since Claude can interpret screenshots from computers connected to the internet, it’s possible that it may be exposed to content that includes prompt injection attacks.
Source-stated
Anthropic says Claude’s computer use remains slow and error-prone, and that it cannot yet attempt some routine computer actions.
Claude’s computer use remains slow and often error-prone. There are many actions that people routinely do with computers (dragging, zooming, and so on) that Claude can’t yet attempt.
Source and contributor context
Anthropic · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Anthropic
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Cursor
Cursor makes an AI coding environment and publishes evaluations of agents working on software tasks.
Editorial provenance
Version 1. Publication ID: 5bd455e9-b55e-18b3-67ea-14512259fc9f.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T18:45:01.823Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 919985f377d131951d639962f0261984ce2064250d7f8962beca431f8206ede8. Source text hash: 8763b8970010b226501c1447a67705573d411138364b1b957b43e0cbe4c0fc0c.
Revision and source-check history
Version 1 · 2026-11-16T18:48:25.243Z
Anthropic describes Claude’s computer-use capability
’s article, published October 22, 2024, describes using a computer through screenshots, movements, clicks and keyboard input. Anthropic argues that this could let AI systems use ordinary software rather than relying on tools specially designed for them. The article’s central idea is to make the model fit the computer environments people already use; Anthropic presents computer use as bringing Claude’s existing abilities to computer tasks, not fundamentally increasing those abilities. Anthropic says Claude learned to interpret screens and act on them, and reports that it generalized from training on simple software. On OSWorld, Anthropic reports a result below human performance but above the next-best AI model in its category. The evaluation does not establish how reliably Claude handles everyday work. The article warns that malicious instructions in screen content could divert Claude from a user’s intent, and that users could misuse computer actions. Anthropic says computer use remains slow and error-prone, with some ordinary actions unsupported. It expects the capability to improve and says researchers will work with safety teams at every stage. How well safeguards prevent unintended actions remains an open question.
Original source
· Article
Source details
Original source date: Oct 22, 2024 (2024-10-22T00:00:00.000Z).
Date basis: Date stated on the publisher’s page.
Added to Escape Meridian: 2026-11-16T18:48:25.243Z. First observed: 2026-10-22T08:23:14.075Z. Retrieved: 2026-10-29T06:31:36.490Z.
Limits
Text as retrieved on October 29, 2026; originally published October 22, 2024. ’s OSWorld result is an evaluation score; the article does not establish how reliably performs everyday work.
Open question
How well will safeguards prevent screen content or user requests from leading to take unintended actions?
Source-stated
Anthropic says Claude can operate a computer through screen-based actions when given the appropriate software setup.
follow a user’s commands to move a cursor around their computer’s screen, click on relevant locations, and input information via a virtual keyboard
Source-stated
Anthropic argues that computer use could let AI systems work with ordinary software rather than relying on bespoke tools.
The next frontier is computer use: AI models that don’t have to interact via bespoke tools, but that instead are empowered to use essentially any piece of software as instructed.
Source-stated
Anthropic reports Claude scored below human performance but above the next-best AI model on OSWorld.
On one evaluation created to test developers’ attempts to have models use computers, OSWorld, Claude currently gets 14.9%. That’s nowhere near human-level skill (which is generally 70-75%), but it’s far higher than the 7.7% obtained by the next-best AI model in the same category.
Source-stated
Anthropic identifies prompt injection as a risk when Claude interprets screenshots from internet-connected computers.
Since Claude can interpret screenshots from computers connected to the internet, it’s possible that it may be exposed to content that includes prompt injection attacks.
Source-stated
Anthropic says Claude’s computer use remains slow and error-prone, and that it cannot yet attempt some routine computer actions.
Claude’s computer use remains slow and often error-prone. There are many actions that people routinely do with computers (dragging, zooming, and so on) that Claude can’t yet attempt.
Source and contributor context
Anthropic · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Anthropic
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Cursor
Cursor makes an AI coding environment and publishes evaluations of agents working on software tasks.
Editorial provenance
Version 1. Publication ID: 5bd455e9-b55e-18b3-67ea-14512259fc9f.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T18:45:01.823Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 919985f377d131951d639962f0261984ce2064250d7f8962beca431f8206ede8. Source text hash: 8763b8970010b226501c1447a67705573d411138364b1b957b43e0cbe4c0fc0c.
uncertain · 2026-11-16T18:45:01.823Z · anthropic/claude-opus-5.5
Publication: 5bd455e9-b55e-18b3-67ea-14512259fc9f
attribution — pass: This is a single Anthropic article, and the reading attributes its descriptions, results, warnings and expectations to Anthropic throughout.
Below, we share some insights from the research that went into developing computer use models—and into making them safer.
summary — uncertain: The reading names OSWorld without saying what it is, and no definition is supplied. A general reader cannot tell that it is a test of how well models use computers. The reading also drops the scores (14.9% for Claude, 7.7% for the next-best model, about 70-75% for humans), which removes the source's own sense of scale. Smallest repair: describe OSWorld as an evaluation built to test models using computers, and give those figures. Optionally, also note Anthropic's argument for releasing computer use now, while the model remains at AI Safety Level 2. The rest of the reading is clear and faithful.
On one evaluation created to test developers’ attempts to have models use computers, OSWorld, Claude currently gets 14.9%.
date — pass: The stated date, October 22, 2024, matches originalPublished at day precision. Anthropic's forecasts, such as its expectation that the capability will improve, are attributed rather than presented as outcomes. The 'Update Aug 28, 2025' banner is a later page addition, but the reading neither uses nor relies on it. The page was edited after publication, but every cited passage is plainly part of the original 2024 article. This is a publisher page, so no archive disclosure is needed.
support — pass: Each claim matches its excerpt. The caveat that the evaluation does not establish everyday reliability is appropriately cautious. The question about screen content or user requests leading to unintended actions follows from the source's warnings about prompt injection and misuse. Although the page was later edited, the cited passages are plainly original article text.
As with any AI capability, there’s also the potential for users to intentionally misuse Claude’s computer skills.
media — pass: No media was selected and no verified candidates were supplied, so nothing invalid was chosen.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeCursor makes an AI coding environment and publishes evaluations of agents working on software tasks.
Explore CursorAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeCursor makes an AI coding environment and publishes evaluations of agents working on software tasks.
Explore CursorAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore Claude