Anthropic has released Claude Opus 4.6, the first Opus-class model featuring a 1 million token context window, delivering significant improvements in agentic workflows and reasoning tasks.
- The model achieves a 68.8% score on ARC AGI 2, nearly doubling Opus 4.5's result of 37.6%.
- It leads in BrowseComp with 84.0%, up from Opus 4.5's 67.8%, and reaches 91.9% on τ2-bench Retail.
- SWE-bench Verified scores are near-parity at 80.8%, slightly down from Opus 4.5's 80.9%.
- A regression is noted in MCP Atlas, dropping from 62.3% to 59.5% for scaled tool use.
The release positions Opus 4.6 as a strong option for research, financial analysis, and knowledge work agents due to its enhanced long-horizon capabilities.