Microsoft's partnership with Anthropic allows Claude Sonnet 4.6 and Opus 4.7 to be used directly inside Copilot Studio as a generative model. This is critical for enterprise customers who find Claude's instruction-following (e.g. for strict JSON schema output) to be superior for specific high-stakes automation. Claude acts as a peer to GPT models, and can be toggled per-environment in the Power Platform Admin Center (PPAC).
Claude is particularly favored for 'Deep Context' RAG because of its 200K–1M token window (Opus 4.7 in beta) and its record-breaking performance in Needle In A Haystack retrieval tests — 98.5% visual acuity benchmark. When an agent needs to analyze a 1,000-page regulatory manual, Claude Opus 4.7 is often the preferred reasoning engine.