Claude 3.7 Sonnet Extended Thinking Release

TL;DR

Anthropic has introduced Claude 3.7 Sonnet, a model that allows users to toggle an "extended thinking mode" to increase the cognitive effort spent on complex tasks. This release is significant because it introduces visible raw thought processes and "action scaling," improving the model's performance in agentic tasks and complex reasoning.

Extended Thinking and Visible Thought Processes

Claude 3.7 Sonnet allows users to direct the model to think more deeply about difficult questions by enabling extended thinking mode. This is not a separate model but the same model allocating more time and effort to reach an answer. Developers can further control this via a "thinking budget."

Benefits of Visibility

Anthropic has made the raw thought process visible to users to achieve several goals:

• Trust: Users can observe and verify the reasoning steps used to reach an answer.
• Alignment: Researchers can identify contradictions between internal thoughts and external responses to detect behaviors like deception.
• Insight: The process reveals reasoning patterns similar to human experts in math and physics, including exploring multiple angles and double-checking answers.

Limitations and Risks

Revealing the thought process introduces specific challenges:

• Faithfulness: Anthropic notes that English-language words in the thought process may not fully represent the model's internal state, and models often make decisions based on factors not explicitly discussed in their thinking.
• Character and Tone: Because standard character training was not applied to the thought process, the internal reasoning is more detached and less personal than the final output.
• Security: Visible thinking could potentially be used by malicious actors to develop better jailbreak strategies or incentivize the model to hide certain thoughts during training.

Agentic Capabilities and Action Scaling

Claude 3.7 Sonnet introduces "action scaling," which improves the model's ability to iteratively call functions and respond to environmental changes to complete open-ended tasks.

Computer Use and OSWorld

The model can issue virtual mouse clicks and keyboard presses to solve tasks. In the OSWorld evaluation, which measures multimodal AI agents, Claude 3.7 Sonnet's performance gap over its predecessor grows as the model is given more interaction steps, indicating better long-term task execution.

Pokémon Red Benchmark

To test agentic capabilities, Anthropic equipped Claude 3.7 Sonnet with basic memory and screen pixel input to play Pokémon Red. While previous Sonnet models failed early (with Claude 3.0 Sonnet unable to leave the starting town), Claude 3.7 Sonnet successfully battled three Gym Leaders and won their Badges by questioning assumptions and trying multiple strategies.

Test-Time Compute Scaling

Claude 3.7 Sonnet's performance scales with the amount of computation used during the inference phase (test-time).

Serial Scaling

The model utilizes "serial test-time compute," meaning it performs sequential reasoning steps before outputting a final answer. Accuracy on mathematical questions improves logarithmically relative to the number of thinking tokens sampled.

Parallel Scaling

Anthropic researchers experimented with "parallel test-time compute," where multiple independent thought processes are sampled and the best is selected via majority voting or a learned scoring function. In the GPQA evaluation (biology, chemistry, and physics), this method—using 256 independent samples and a 64k-token thinking budget—resulted in a GPQA score of 84.8%, including a physics subscore of 96.5%.

Safety and Security Frameworks

AI Safety Level (ASL-2)

Anthropic confirmed that Claude 3.7 Sonnet meets the ASL-2 safety standard. While the model showed some "uplift" in assisting participants with tasks related to Chemical, Biological, Radiological, and Nuclear (CBRN) weapons, all attempts contained critical failures that prevented successful end-to-end completion.

Thought Process Safeguards

To prevent the exposure of harmful content, Anthropic implements encryption for the thought process. If the thinking contains high-risk topics (e.g., cyber attacks or dangerous weapons), the user will see the message: "the rest of the thought process is not available for this response."

Prompt Injection Defense

For computer use capabilities, Anthropic improved defenses against prompt injection attacks. By combining new training, a revised system prompt, and a dedicated classifier, the model now prevents these attacks 88% of the time, an increase from 74%.