Claude Sonnet 4.6 hits near-human 72.5% on OSWorld

Claude Sonnet 4.6 is a category-defining agentic model that renders previous "computer use" capabilities obsolete and puts massive pressure on the flagship tiers of competitors.

• –The leap from 28% to 72.5% on OSWorld-Verified in a single year signals the arrival of reliable, autonomous agent workflows for complex desktop tasks.
• –A 79.6% score on SWE-bench Verified cements its status as the premier model for autonomous software engineering, outperforming the previous flagship Opus 4.5.
• –The 1 million token context window (now generally available) enables seamless reasoning across massive codebases and document sets without the overhead of complex RAG pipelines.
• –Pricing remains at the standard Sonnet rate, offering a massive intelligence-to-cost ratio that undercuts GPT-5.2 on both performance and efficiency.