Opus 4.6, Now Live in :Harvey:
Announcing Claude Opus 4.6 in Harvey.
Today, we are making Claude Opus 4.6 available in Harvey. Opus 4.6 is Anthropic's latest frontier model, extending the strengths of Opus 4.5 in agentic problem-solving while delivering meaningful improvements across complex legal workflows.
Early access evaluations point to gains in substantive accuracy, source grounding, and the ability to handle multi-step legal tasks with nuance and precision. We expect Opus 4.6 to excel in research-intensive work, due diligence, and complex transactional matters where getting the details right is essential.
On our BigLaw Bench evaluation suite, Opus 4.6 scored 90.2% — the highest score yet for the Claude family of models. With 40% of tasks receiving perfect scores, the model demonstrates strong legal reasoning capability across both litigation and transactional practice areas. Deal management, risk assessment, and corporate strategy tasks were particular standouts.
During early access testing, our Applied Legal Research team highlighted several strengths. Outputs have high substantive accuracy, with improved analytical depth on complex, multi-faceted legal questions. The model has also resolved stylistic issues from earlier versions: final answers are cleaner, without the preamble text that sometimes appeared in prior Claude models. The model's source grounding is notably thorough — in one complex research query, it generated over 120 inline citations, each tied to specific passages in the source material.
“Opus 4.6 reflects remarkable progress in agentic capabilities. We’ve seen meaningful gains in multi-step reasoning and task completion. This release is another step toward AI that can work alongside lawyers on their most challenging matters, and Harvey clients are once again among the first to benefit.”
Niko Grupen
Head of Applied Research at Harvey
We also observed an interesting pattern in how the model calibrates its responses under uncertainty. When Opus 4.6 is confident, it delivers precise, high-quality outputs of appropriate length. When less certain, it demonstrates a tendency to over-contextualize responses, producing unnecessary tokens along the way. We’re continuing to refine how we deploy extended thinking for legal workflows that need clear, succinct output.
We'll be rolling out Opus 4.6 in the model selector to eligible clients in the US and EU in the coming days. AU availability will be announced shortly.
