AI That Thinks Before It Speaks? Claude 3.7 Sonnet Debuts Hybrid Reasoning

Anthropic has rolled out Claude 3.7 Sonnet, a new version of its AI model that introduces an optional “extended thinking” mode, giving users more control over how the model processes complex tasks.

Anthropic has released Claude 3.7 Sonnet, an updated AI model that introduces hybrid reasoning, allowing users to choose between rapid responses and extended thinking.

This new approach aims to improve problem-solving by enabling the model to take additional time to evaluate and refine its responses when needed.

In standard mode, Claude 3.7 Sonnet functions as an iterative improvement over its predecessor, Claude 3.5 Sonnet. In extended thinking mode, however, the model engages in self-reflection before producing an answer.

Michael Gerstenhaber, Anthropic’s product lead for AI, described the feature as a way for users to balance response time and reasoning depth. “The [user] has a lot of control over the behavior—how long it thinks, and can trade reasoning and intelligence with time and budget” he said, as reported by WIRED.

This process is intended to enhance accuracy in areas requiring multi-step reasoning, such as mathematics, physics, coding, and instruction-following. Users interacting with the model via API can specify a limit on how many tokens it uses to “think,” allowing control over response speed and cost.

WIRED notes that the update also includes a “scratchpad” feature, which makes the model’s reasoning steps visible as it processes a request. This approach is similar to techniques used in other AI models, such as DeepSeek’s.

Anthropic states that Claude 3.7 Sonnet has been optimized for real-world applications rather than competition-style benchmarks. Early testing suggests that the extended reasoning mode improves performance on tasks involving complex codebases, advanced tool use, and full-stack software development.

The model has been evaluated on SWE-bench Verified and TAU-bench, where it reportedly outperforms previous versions, as reported by Anthropic.

CNN Notes that Anthropic has kept the pricing for Claude 3.7 the same as previous models, positioning it below OpenAI’s latest offerings. It costs $3 per million input tokens and $15 per million output tokens, compared to OpenAI’s $15 and $60 for similar tiers.