Anthropic’s already impressive Claude 3.5 Sonnet gains a significant performance boost on Tuesday as the generative AI startup rolls out an enhanced and updated version of the model alongside the new, lightweight Claude 3.5 Haiku. The Sonnet update includes a public beta feature that gives the AI basic control over the computer it’s running on.
Claude 3.5 Sonnet was already a performance leader when it comes to coding tasks, but the new version shows significant across-the-board improvements over its predecessor and steadily outperforms both Gemini 1.5 and GPT-4o on a variety of industry benchmarks. Gemini 1.5 Pro was the only model to best the new 3.5 Sonnet on any test, and did so on the MATH benchmark.
The new 3.5 Haiku is no slouch, either, despite its small size. Scheduled to be released later this month, 3.5 Haiku outperforms Claude 3.0 Opus, the company’s largest last generation model. Like its larger version, the new Haiku is exceedingly proficient at coding tasks, scoring 40.6% on the SWE-bench Verified — higher than both GPT-40 and the original 3.5 Sonnet.
Even more impressive, the new Claude 3.5 Sonnet can now interact with desktop apps via the “Computer Use” API. The AI can generate the necessary keystrokes, mouse clicks, and movements needed to emulate the human user. The company is quick to point out that the system is currently quite experimental and prone to errors. The underlying purpose of the public beta release is to elicit feedback from developers to rapidly improve the API’s performance.
“We trained Claude to see what’s happening on a screen and then use the software tools available to carry out tasks,” Anthropic wrote in a blog post. “When a developer tasks Claude with using a piece of computer software and gives it the necessary access, Claude looks at screenshots of what’s visible to the user, then counts how many pixels vertically or horizontally it needs to move a cursor in order to click in the correct place.”
It’s an AI agent, essentially. That is, its an AI that can automate other software processes, whether that’s generating and qualifying marketing leads, uncovering patterns and trends in medical data, or simply navigating to a specific website and filling out a form you need. Think of them as a more advanced version of existing Robotic Process Automation systems.
The company cites Asana, Canva, Cognition, DoorDash, Replit, and The Browser Company as early adopters of the new feature. Replit, for example, is using Computer Control to “develop a key feature that evaluates apps as they’re being built for their Replit Agent product,” per the announcement.
There’s no need to worry about the AI going all Skynet on us (yet), as Anthropic explains. “Humans remain in control by providing specific prompts that direct Claude’s actions, like ‘use data from my computer and online to fill out this form,’” an Anthropic spokesperson told TechCrunch. “People enable access and limit access as needed. Claude breaks down the user’s prompts into computer commands (e.g., moving the cursor, clicking, typing) to accomplish that specific task.”
Anthropic also concedes that Computer Control could be misused to generate spam, spread misinformation, or commit fraud. In response, the company has developed new classifiers that identify when the API is being used and whether that use is “causing harm.”
