Claude 3.5 New Features Breakdown: Computer Control and Coding Capabilities Get a Major Upgrade

Anthropic recently delivered a major upgrade to its Claude 3.5 Sonnet model, adding the ability to control a computer and achieving another leap in coding performance. The update also introduces the all-new Claude 3.5 Haiku model, further diversifying its product offerings. Here’s a look at the standout features.

Claude 3.5 Sonnet Gains Computer Control

The most eye-catching addition to the new Claude 3.5 Sonnet is its ability to operate a computer. Anthropic built a dedicated API that lets the model perceive a computer interface and interact with it much like a human would. Developers can integrate the API to have Claude perform tasks such as moving the cursor, clicking buttons, and filling out forms.

In the OSWorld benchmark, Claude 3.5 Sonnet scored 14.9% in screenshot-only mode, significantly outperforming other AI systems. While the model still faces challenges with scrolling or dragging, well-known companies like Asana and Replit are already testing the feature. This capability opens up new possibilities for automating repetitive workflows.

Significant Coding Improvements and Performance Optimization

The updated Claude 3.5 Sonnet delivers a major leap in coding performance. On the SWE-bench Verified test, its score jumped from the previous 33.4% to 49.0%, surpassing reasoning models including o1-preview. Early client feedback shows that GitLab saw a 10% improvement in reasoning for DevSecOps tasks without any increase in latency when using the model.

Claude 3.5 Haiku, the fastest new model, also performs strongly on coding tasks. It achieved a score of 40.6% on SWE-bench Verified, exceeding many publicly available models while maintaining the same cost and speed as the previous Haiku generation. These advances make Claude more reliable when tackling complex software engineering tasks.

Additional Performance Gains and Model Options

Beyond the core updates, the new Claude 3.5 Sonnet also showed improvements across several benchmarks. On the TAU-bench evaluation, its score in the retail domain rose by 6.6 percentage points to 69.2%. Claude 3.5 Haiku retains the advantages of low cost and high speed, making it ideal for user-facing products or scenarios requiring fast responses.

Claude 3.5 Sonnet is now available to all users, and developers can access it through the Anthropic API or platforms like Amazon Bedrock. This upgrade not only strengthens Claude’s leadership in the programming space but also marks a significant step toward AI models that truly understand and interact with the digital world.