Sonnet 4.6, Now Live in :Harvey:

Today, we are making Claude Sonnet 4.6 available in Harvey. Sonnet 4.6 is Anthropic's latest model in the Sonnet family, which bridges strong reasoning with a faster, more efficient architecture. Early access evaluations highlight strengths in instruction-following, practical problem solving, and ideation.

On our BigLaw Bench evaluation suite, Sonnet 4.6 scored 87.6%, with 35% of tasks receiving perfect scores. Model performance was balanced across practice areas, with Sonnet 4.6 receiving the same average score across litigation and transactional tasks.

Our Applied Legal Research team highlighted multiple strengths during early access testing. The model demonstrated strong numerical precision, delivering more accurate specific figures like dollar amounts compared to prior Claude models when explicitly instructed. On ideation-oriented tasks, including trial exhibit preparation and open-ended litigation strategy questions, evaluators noted that Sonnet 4.6's thoroughness was a clear advantage. In Claude Code, our Engineering and Research teams noted that the model excelled at root cause analysis, correctly tracing bugs through code history and integrating parallel tool calls to solve problems that even Opus 4.6 missed.

Evaluators also observed an interesting pattern in how Sonnet 4.6 responds to explicit versus implicit instructions. When given specific direction (e.g. to retrieve exact figures, follow a defined rubric, or produce a structured comparison), the model is thorough and precise. When left to determine scope on its own, it tends to remain at too high a level, providing reasonable summaries, but missing fine-grained details. This gap was most visible on tasks requiring synthesis across multiple documents.

We'll be rolling out Sonnet 4.6 in the model selector to eligible clients in the US and EU in the coming days. AU availability will be announced shortly.