Anthropic just dropped Sonnet 4.6...

Overview

Anthropic's Claude Sonnet 4.6 is positioned as a workhorse model for knowledge workers, featuring improved coding skills, consistency, and instruction following, as well as a 1 million token context window. Benchmarks show Sonnet 4.6 outperforming Sonnet 4.5 in agentic terminal coding (51% to 59%) and tool use (43.8% to 61.3%), even surpassing Opus 4.6 and other models in certain financial analysis tasks.

Key takeaways

Claude Sonnet 4.6 is designed as a workhorse model for knowledge workers, featuring improved coding skills, consistency, and instruction following.
Sonnet 4.6's tool use capabilities have significantly improved, jumping from 43.8 to 61.3 in benchmarks, making it more valuable for real-world applications.
In agentic financial analysis, Sonnet 4.6 outperforms Opus 4.6, Gemini 3 Pro, and GPT 5.2, demonstrating its strength in knowledge work.
Anthropic has deployed Sonnet 4.6 under AI safety level 3, indicating a heightened risk of catastrophic misuse compared to non-AI systems.
Sonnet 4.6 supports adaptive reasoning, allowing users to scale thinking tokens up or down as needed.
Anthropic acknowledges that it's becoming increasingly difficult to determine if models are capable of reaching AI R&D4 and CBRN4 thresholds due to their increasing capabilities.

Chapters

0:00 Introducing Claude Sonnet 4.6: A Powerful Model for Real-World Tasks

Claude Sonnet 4.6 features improved coding skills, consistency, instruction following, and a 1 million token context window.
The model is positioned for real-world tasks, including creating PowerPoints and manipulating Excel work within Claude's co-work environment.
Anthropic emphasizes the importance of safety and security, addressing risks like prompt injection attacks and working to improve resistance.

7:08 Benchmark Performance: Sonnet 4.6 Excels in Agentic Tasks and Tool Use

Sonnet 4.6 shows significant performance jumps compared to Sonnet 4.5 in agentic terminal coding (51% to 59%) and agentic computer use (61% to 72%).
In tool use, Sonnet 4.6 jumps from 43.8 to 61.3, making it more valuable for real-world use cases involving querying and using MCP servers.
Sonnet 4.6 outperforms other models, including Opus 4.6, Gemini 3 Pro, and GPT 5.2, in agentic financial analysis.

11:46 Adaptive Reasoning, Safety Levels, and the Future of Model Capabilities

Sonnet 4.6 outperforms Sonnet 4.5 on Vending Bench, achieving $5,500 in simulated profit after 350 days by investing in capacity early and pivoting to profitability.
Anthropic has deployed Sonnet 4.6 under AI safety level 3, indicating a substantial increase in the risk of catastrophic misuse compared to non-AI baselines.
Anthropic acknowledges that confidently ruling out AI R&D4 and CBRN4 capability thresholds is becoming increasingly difficult due to the model's high levels of capability.

Summary, takeaways, and chapters were generated by AI from the video's transcript and may contain errors. The video belongs to its creator, Matthew Berman.