Claude Opus 4.6

Claude Opus 4.6 positions itself as a high-capability reasoning and agentic work model, building on the Opus line's legacy of deep analytical performance. The release emphasizes a qualitative shift in long-context reliability, supported by a notably higher score on the MRCR v2 multi-needle evaluation compared to its predecessor, which enables more trustworthy performance when reasoning across very large document sets. Its adaptive thinking mechanism introduces four effort levels — low, medium, high, and max — allowing the model to dynamically calibrate depth based on task complexity, with high as the default. The Compaction API further extends practical usability by automatically summarizing older conversation context server-side, enabling sustained interactions without manual management. For code-heavy workflows, the model demonstrates state-of-the-art performance on Terminal-Bench 2.0 and leads on Humanity's Last Exam, a complex multidisciplinary reasoning evaluation.

Training lineage details are not fully disclosed in the supplied sources, but the model's performance trajectory on GDPval-AA — outperforming the next-best model by roughly 144 Elo points and its own predecessor by 190 points — suggests meaningful post-training refinement aimed at economically valuable knowledge work in finance, legal, and related domains. The model is designed to sustain agentic tasks longer and operate more reliably within large codebases, addressing a common pain point in coding-focused deployments. Its integration within Cowork allows Opus 4.6 to apply planning, research, document creation, and financial analysis capabilities in a multi-step autonomous workflow. Anthropic's own evaluation reporting highlights top-tier results on BrowseComp for hard-to-locate information retrieval, indicating a focus on research-grade reliability alongside everyday task performance.