Anthropic's New AI Model: Long-Form Thought Capabilities

Anthropic Unveils Claude 3.7 Sonnet: A New AI Model with Adaptive Reasoning

Anthropic has announced the release of Claude 3.7 Sonnet, a cutting-edge AI model engineered to process inquiries with a variable degree of cognitive effort, determined by user preference.

This new model is designated by Anthropic as the industry’s inaugural “hybrid AI reasoning model,” integrating both immediate response capabilities and a more deliberate, analytical approach to question answering. Users have the option to activate the model’s reasoning functions, allowing Claude 3.7 Sonnet to dedicate varying amounts of time to contemplation.

Simplifying the AI User Experience

The introduction of Claude 3.7 Sonnet reflects Anthropic’s commitment to streamlining the interaction with its AI offerings. Many current AI chatbots present users with a complex selection of models, differing in both cost and performance. Anthropic aims to eliminate this complexity, envisioning a single model capable of handling all tasks.

The rollout of Claude 3.7 Sonnet to all users and developers begins on Monday, according to Anthropic. However, access to the model’s reasoning features will be exclusive to subscribers of Anthropic’s premium Claude plans.

Users of the free Claude version will receive the standard iteration of Claude 3.7 Sonnet, which Anthropic asserts surpasses the performance of its prior frontier model, Claude 3.5 Sonnet. (Notably, the company bypassed the number 3.6 in the naming sequence.)

Cost and Comparison to Competitors

Claude 3.7 Sonnet is priced at $3 per million input tokens – equivalent to approximately 750,000 words, exceeding the length of the “Lord of the Rings” trilogy – and $15 per million output tokens.

While this represents a higher cost than OpenAI’s o3-mini ($1.10 per 1 million input tokens/$4.40 per 1 million output tokens) and DeepSeek’s R1 (55 cents per 1 million input tokens/$2.19 per 1 million output tokens), it’s important to note that o3-mini and R1 are dedicated reasoning models, unlike the hybrid nature of Claude 3.7 Sonnet.

The Rise of Reasoning Models

Claude 3.7 Sonnet marks Anthropic’s initial foray into AI models capable of “reasoning,” a technique gaining prominence as conventional methods of enhancing AI performance reach their limits.

Reasoning models, such as o3-mini, R1, Google’s Gemini 2.0 Flash Thinking, and xAI’s Grok 3 (Think), allocate additional time and computational resources to problem-solving before providing responses. They decompose complex issues into smaller, manageable steps, generally leading to improved accuracy. This process, while not mirroring human thought, is inspired by deductive reasoning.

Future Development and Transparency

Anthropic intends for Claude to eventually autonomously determine the optimal duration of contemplation required for each question, eliminating the need for user-defined controls. This vision was shared by Dianne Penn, Anthropic’s product and research lead, in an interview with TechCrunch.

“We view reasoning as an inherent capability of a frontier model, seamlessly integrated with others, rather than requiring a separate model,” Anthropic stated in a blog post provided to TechCrunch. “This is akin to how humans don’t employ distinct brains for immediate versus thoughtful responses.”

Anthropic is introducing a “visible scratch pad” feature, allowing users to observe Claude 3.7 Sonnet’s internal planning process. Penn explained to TechCrunch that users will generally have access to the model’s complete thought process, with certain portions potentially redacted for safety and trust considerations.

Optimized for Real-World Applications

Anthropic has specifically optimized Claude’s reasoning modes for practical applications, including complex coding challenges and agentic tasks. Developers utilizing Anthropic’s API can regulate the “budget” allocated to reasoning, balancing speed, cost, and answer quality.

In testing, Claude 3.7 Sonnet achieved 62.3% accuracy on the SWE-Bench test, which measures real-world coding tasks, compared to OpenAI’s o3-mini model’s score of 49.3%. On the TAU-Bench test, assessing AI interaction with simulated users and APIs in a retail environment, Claude 3.7 Sonnet scored 81.2%, while OpenAI’s o1 model achieved 73.5%.

Furthermore, Anthropic reports that Claude 3.7 Sonnet will decline to answer inappropriate questions less frequently than its predecessors, demonstrating an enhanced ability to differentiate between harmful and harmless prompts. The company has reduced unnecessary refusals by 45% compared to Claude 3.5 Sonnet.

Introducing Claude Code

Alongside Claude 3.7 Sonnet, Anthropic is launching Claude Code, an agentic coding tool currently available as a research preview. This tool enables developers to execute specific tasks through Claude directly from their terminal.

During a demonstration, Anthropic personnel showcased Claude Code’s ability to analyze a coding project with a simple command, such as, “Explain this project structure.” Developers can utilize plain English commands to modify codebases, with Claude Code providing descriptions of its edits and even testing the project for errors or pushing changes to a GitHub repository.

Access to Claude Code will initially be granted on a “first come, first serve” basis to a limited number of users, as confirmed by an Anthropic spokesperson to TechCrunch.

A Competitive Landscape

Anthropic’s release of Claude 3.7 Sonnet occurs within a rapidly evolving AI landscape, characterized by frequent model releases. While Anthropic has traditionally prioritized a methodical, safety-conscious approach, the company now aims to establish itself as a leader in the field.

However, the longevity of this leadership remains uncertain, as OpenAI is reportedly developing its own hybrid AI model, anticipated to be released within “months,” according to CEO Sam Altman.