Claude 3.7 Sonnet introduces “extended thinking mode” to solve complex problems

Claude 3.7 Sonnet, the latest version of Anthropic's AI assistant, incorporates an innovative feature called "extended thinking mode" that transforms its ability to solve complex problems. This capability allows the model to dedicate more time and computational resources to difficult questions, similar to when humans apply greater mental effort to challenging tasks.

The innovation isn't just about improved performance, but Anthropic's decision to make the model's reasoning process visible. This offers several benefits: it builds greater trust by allowing verification of how answers are reached, facilitates system alignment by identifying potential problematic behaviors, and makes it fascinating to observe how Claude approaches mathematical or physical problems in ways similar to human experts.

The extended thinking mode has significantly improved Claude's performance in standard evaluations and unexpected tasks. For example, in tests like OSWorld, which measures multimodal capabilities, or even playing Pokémon Red, where it vastly outperformed previous versions by maintaining focus and achieving complex objectives.

Anthropic has also experimented with improvements through parallel computing, sampling multiple independent thought processes and selecting the best one, which has produced outstanding results in complex scientific evaluations.

These enhanced capabilities come with reinforced security measures, including encryption of certain potentially harmful thought processes and protections against "prompt injection" attacks. Claude 3.7 Sonnet is now available both through Claude.ai and via API.