Anthropic Unveils Claude 3.5 Sonnet and Haiku Models with New Computer Use Feature
Anthropic, a prominent AI company, has announced the release of its upgraded Claude 3.5 Sonnet and a new model, Claude 3.5 Haiku, according to anthropic.com. The Claude 3.5 Sonnet showcases significant improvements, particularly in coding, while Claude 3.5 Haiku matches the performance of Claude 3 Opus, their previous largest model, on several benchmarks.
Advancements in AI Capabilities
The Claude 3.5 Sonnet model exhibits comprehensive upgrades, particularly in coding tasks, where it has led the field. It enhances performance on industry benchmarks such as SWE-bench Verified, improving from 33.4% to 49.0%, surpassing other publicly available models. The model also shows improvements in agentic tool use tasks, with significant gains in both retail and airline domains.
Similarly, Claude 3.5 Haiku is positioned as a cost-effective and fast alternative, surpassing the Claude 3 Opus in various intelligence benchmarks. It is particularly strong in coding tasks, scoring 40.6% on SWE-bench Verified, which outperforms several state-of-the-art models.
Introducing Computer Use in Public Beta
Anthropic is also pioneering a new capability called 'computer use' in public beta. This feature allows developers to instruct Claude to interact with computers similarly to humans, involving actions like moving cursors and clicking buttons. While currently experimental, it opens new possibilities for automating complex tasks requiring multiple steps. Companies like Replit and The Browser Company are already exploring these capabilities for various applications.
The computer use feature is available via the Anthropic API, Amazon Bedrock, and Google Cloud’s Vertex AI. It provides a novel approach for developers to automate repetitive processes and conduct open-ended tasks, although it currently faces challenges with basic actions like scrolling and zooming.
Ensuring Responsible Deployment
To ensure the safe deployment of these new capabilities, Anthropic has partnered with the US AI Safety Institute and the UK Safety Institute for pre-deployment testing. They have also developed classifiers to detect misuse of the computer use feature, aiming to mitigate risks such as spam and misinformation.
Anthropic is committed to continuous improvement of these models and features, anticipating rapid advancements in the coming months. The release of Claude 3.5 Haiku is scheduled for later this month, initially as a text-only model with plans for image input capabilities.
Looking Forward
These developments are expected to enhance the way users interact with AI, offering new possibilities for automation and personalization in various domains. Anthropic invites feedback from developers to refine these capabilities further.