Claude-Opus-4.6

Claude Opus 4.6 is Anthropic's latest flagship model, building on its predecessor with meaningful gains in software engineering, agentic task execution, and extended reasoning workflows. The model introduces Adaptive Thinking and a Context Compaction API designed to combat the gradual loss of useful context that typically degrades performance in long-running autonomous agents. It handles knowledge work ranging from financial analysis and document creation to multi-step research pipelines, and its architecture supports extended engagement with large codebases through improved planning, code review, and debugging capabilities that allow it to catch its own mistakes. The model achieves state-of-the-art results on multiple industry evaluations, including the top score on Terminal-Bench 2.0 for agentic coding and leading performance on Humanity's Last Exam for complex multidisciplinary reasoning.

The training and post-training process incorporated interpretability tools—including activation oracles, attribution graphs, and sparse autoencoder features—as practical methods for investigating model behavior during development. Safety evaluations reported broadly improved capabilities compared to prior generations with a comparably low rate of misaligned behavior overall, though specific areas like sabotage concealment and overly agentic behavior in computer-use settings received additional scrutiny during the alignment assessment. On GDPval-AA, an evaluation of performance across finance, legal, and other economically valuable knowledge domains, Opus 4.6 outperforms the next-best frontier model by a notable margin and shows a substantial jump over its own predecessor. These improvements position the model well for extended, autonomous workflows where sustained reliability and error recovery matter over long execution cycles.