NEW - Anthropic Updated Claude Models & Computer Use Agents!!

Anthropic’s latest release pairs two upgraded Claude models with a new “computer use” capability that lets Claude interact with a user’s computer directly—turning natural-language requests into mouse clicks, keyboard input, and browser actions. The practical impact is straightforward: instead of relying on separate tools or search workflows, Claude can operate the interface itself, which makes agent-style automation feel much closer to “do the task for me” than “generate code or instructions I must run.”

The model lineup starts with an upgraded Claude 3.5 Sonnet, available immediately across Anthropic’s platform and also through common deployment routes like Google Cloud Vertex and Amazon Bedrock. Benchmarks compare the new Sonnet against the prior Claude 3.5 Sonnet and several competitors (including GPT-4o, GPT-4o-mini, Gemini 1.5 Pro, and Gemini 1.5 Flash) using zero-shot evaluations. The new Sonnet improves across nearly all tests, with one exception where Gemini 1.5 Pro remains ahead. The most notable gains show up in coding and agentic tool use—especially SWE-bench Verified, where performance rises from 33.4% to 49%. The transcript frames this as a meaningful step for developers and coding assistants, particularly for workflows associated with agentic coding and tool-using agents.

A second announcement brings Claude 3.5 Haiku, which is not available immediately but is expected later this month. Haiku has historically appealed for being fast and inexpensive while still delivering strong results, and the new version is positioned as even more capable—reportedly surpassing Claude 3 Opus on many tasks. In SWE-bench scoring, the updated Haiku lands higher than the previous Sonnet version, reinforcing the idea that it’s optimized for high-throughput “agent calls” where speed and cost matter. The first release is text-only, with image input expected to follow.

The biggest shift, though, is Anthropic’s computer use API. Rather than calling external tools for search or navigation, Claude can take actions directly in the user’s environment—opening a browser, performing searches, clicking through pages, and entering text. The transcript highlights OSWorld as a benchmark used to evaluate this kind of interface-level control, and suggests that pairing computer use with stronger models (like the newer Sonnet) should improve real-world agent performance.

Still, there’s a clear caution: giving an AI control over a computer introduces real risk. During Anthropic’s own demonstrations, the system reportedly produced “amusing errors,” including stopping recording and wandering into unrelated browsing (such as searching for Yellowstone National Park photos). The transcript also notes that commenters previously raised the idea of using a dedicated machine for such experiments.

In demos, Claude receives a screenshot of what’s on screen, then responds with concrete actions—mouse movement, clicks, and typing—enabling it to edit documents and complete multi-step tasks. The transcript’s takeaway is that Sonnet may serve as an “orchestrator” for complex work, while Haiku could handle rapid, repeated agent steps, and computer use could unify these into end-to-end automation—assuming users implement safeguards and retain the ability to stop or constrain the system quickly.