Anthropic Launches Claude 3.7 Sonnet, Combining Fast Responses with Deep Reasoning Abilities

Anthropic has introduced its latest AI model, Claude 3.7 Sonnet, which is positioned as the first "hybrid reasoning model" in the market. This model can seamlessly switch between the instant response mode typical of traditional AI systems and a visible, step-by-step thinking process, simulating the human ability to transition from quick reactions to deep contemplation.

Claude 3.7 Sonnet integrates immediate responses with extended thinking modes, a feature that sets it apart from competitors who typically separate these functionalities. Additionally, the model excels in coding tasks, achieving state-of-the-art results on real-world software benchmarks such as SWE-bench Verified.

To complement the launch of this model, Anthropic has also released Claude Code, an agent-based terminal tool capable of reading code repositories, editing files, and even pushing changes to GitHub repositories. This tool further enhances the application potential of Claude 3.7 Sonnet in the development field.

In terms of pricing, Claude 3.7 Sonnet maintains the same cost structure as its predecessor: $3 per million input tokens and $15 per million output tokens (including thinking tokens). This pricing strategy provides users with flexible options while ensuring the model remains economically viable.

Claude 3.7 Sonnet introduces a new approach to handling complex problems by integrating rapid response capabilities with deep reasoning functions within a single system. This design allows for greater flexibility and efficiency in task processing, avoiding the need to split these abilities across different models.

For API users, Anthropic offers granular control options. Users can specify the maximum number of tokens the model should use for thinking before responding (up to the 128K output limit), allowing for customizable trade-offs between quality and speed. This helps organizations find a balance between performance requirements and cost considerations.

Claude 3.7 Sonnet's performance in coding tasks is particularly noteworthy. It achieved outstanding results in benchmark tests like SWE-bench Verified and TAU-bench, assessing the model's ability to solve real-world software problems and handle complex user-tool interactions.

To further enrich its product ecosystem, Anthropic has launched Claude Code, an agent-based coding assistant. Currently in a limited research preview phase, this tool can already perform tasks such as searching and reading code repositories, editing files, writing and running tests, and submitting changes to GitHub repositories. Although still in its early stages, Claude Code has already demonstrated significant improvements in productivity.

Notably, the extended thinking functionality of Claude 3.7 Sonnet is available across all paid plans, including Pro, Team, and Enterprise plans, as well as through Anthropic API, Amazon Bedrock, and Google Cloud’s Vertex AI platforms. However, free-tier users will not have access to these extended thinking features.

Regarding security, Claude 3.7 Sonnet has been improved to more accurately distinguish harmful requests from benign ones, reducing unnecessary rejections by 45% compared to its predecessor. This enhancement contributes to a better user experience and increased trust in the model.