Anthropic announced last Monday Claude 3.7 Sonnet, which they describe as their "most intelligent model to date and the first hybrid reasoning model on the market" along with Claude Code, an agentic coding tool that allows developers to automate complex tasks directly from their terminal.
Claude 3.7 Sonnet: Significant Improvements in Coding and Front-End Web Development
Anthropic adopts a unified approach to reasoning by combining speed and depth of analysis. Unlike other models that develop separate solutions for instant reasoning and deep reflection, Claude 3.7 Sonnet integrates these two capabilities into a single system.
The model offers two operating modes:
Standard, which enhances the performance of Claude 3.5 Sonnet with more accurate and natural responses.
Extended reflection, step-by-step, visible to the user where the AI takes time to analyze a problem before formulating a response, thus improving its performance in mathematics, coding, and following complex instructions.
One of the key innovations of Claude 3.7 Sonnet is the ability for Anthropic API users to set a reflection budget in terms of token count. This granular control allows finding a balance between response quality and operational costs.
A Strategic Positioning Between Performance and Security
Anthropic emphasizes the security and reliability of its models. Extensive testing and collaborations with external experts have been conducted to ensure that the model effectively distinguishes between harmful and benign requests, reducing unjustified refusals by 45% compared to its predecessor. The company has published a detailed system card of its evaluation methods, covering risks of rapid injection attacks and strategies implemented to counter them.
Claude Code: Towards Advanced Software Development Automation
With Claude Code, Anthropic aims to transform development practices by offering an AI assistant capable of searching and modifying code, running tests, and even pushing changes to GitHub.
Initial tests show a significant reduction in the time required to accomplish certain tasks, sometimes by more than 45 minutes. This improvement could profoundly change the daily life of developers, especially in areas like debugging, large-scale refactoring, and test-driven development.
Availability
Claude 3.7 Sonnet is available on all Claude plans, including Free, Pro, Team, and Enterprise, as well as via the Anthropic API, Amazon Bedrock, and Google Cloud's Vertex AI. However, the extended reflection mode is reserved for users of the paid plans. Pricing remains the same as previous versions: $3 per million input tokens and $15 per million output tokens, including tokens used for reflection.
On the other hand, Claude Code is currently available as a limited research preview. Based on received feedback, Anthropic will improve "the reliability of tool calls, add support for long-duration commands, improve rendering in the app, and expand Claude's understanding of its capabilities."