Microsoft blends rival LLMs
What happened
Microsoft upgraded Copilot to orchestrate multiple LLMs—using GPT for drafting, Claude for critique—and shipped Copilot Cowork for long‑running, multi‑step automation and a Researcher agent. The move signals multi‑model orchestration as a platform pattern and sets expectations for role‑based model selection in developer tools. (thenewstack.io) (theverge.com)
Why it matters
Copilot Cowork moved into Microsoft’s Frontier early‑access program on March 30, 2026. (microsoft.com) Microsoft describes Cowork as creating a plan, reasoning across Outlook, Teams, Excel and other work content, and carrying multi‑step tasks forward with visible progress and built‑in “skills” drawn from Anthropic and Microsoft toolchains. (microsoft.com) Researcher gained two named multi‑model capabilities—Critique and Council—that separate generation from evaluation and are implemented using combinations of models from Frontier partners, including Anthropic and OpenAI. (techcommunity.microsoft.com) Microsoft’s support docs say Critique runs a second reasoning pass (the “Critique model”) that applies Claude to strengthen structure, prioritize reputable sources, and add citation grounding, and that Critique becomes the default when Anthropic is enabled. (support.microsoft.com) Copilot Studio exposes a drop‑down model selector and multi‑agent tools so teams can mix Anthropic Sonnet 4 / Opus 4.1, OpenAI models, and entries from the Azure Model Catalog when building agents; Anthropic model choice began rolling out on Sept. 24, 2025 with admin opt‑in via Frontier. (microsoft.com) Microsoft announced a new Microsoft 365 E7 enterprise tier priced at $99 per user launching May 1, and a separate Agent 365 product priced at $15, with Microsoft saying E7 is roughly 65% more expensive than the current $60 E5 plan. (geekwire.com) Independent coverage and analysis flagged governance tradeoffs: Anthropic’s models are hosted outside Microsoft‑managed environments and commentators warned that AWS hosting can add latency, egress costs and data‑sovereignty complications that drive the need for model‑aware routing and monitoring. (microsoft.com) Microsoft and reporting outlets emphasize that Cowork runs inside Microsoft 365 security and compliance boundaries and that actions and outputs are auditable by default, and Microsoft cited Capital Group as an early access customer using Cowork in production scenarios. (microsoft.com)
Key numbers
- (thenewstack.io) (theverge.com) Copilot Cowork moved into Microsoft’s Frontier early‑access program on March 30, 2026.
- 24, 2025 with admin opt‑in via Frontier.
- (microsoft.com) Microsoft announced a new Microsoft 365 E7 enterprise tier priced at $99 per user launching May 1, and a separate Agent 365 product priced at $15, with Microsoft saying E7 is roughly 65% more expensive than the current $60 E5 plan.
What happens next
- (microsoft.com) Microsoft describes Cowork as creating a plan, reasoning across Outlook, Teams, Excel and other work content, and carrying multi‑step tasks forward with visible progress and built‑in “skills” drawn from Anthropic and Microsoft toolchains.
- (microsoft.com) Microsoft announced a new Microsoft 365 E7 enterprise tier priced at $99 per user launching May 1, and a separate Agent 365 product priced at $15, with Microsoft saying E7 is roughly 65% more expensive than the current $60 E5 plan.
Quick answers
What happened in Microsoft blends rival LLMs?
Microsoft upgraded Copilot to orchestrate multiple LLMs—using GPT for drafting, Claude for critique—and shipped Copilot Cowork for long‑running, multi‑step automation and a Researcher agent. The move signals multi‑model orchestration as a platform pattern and sets expectations for role‑based model selection in developer tools. (thenewstack.io) (theverge.com)
Why does Microsoft blends rival LLMs matter?
Copilot Cowork moved into Microsoft’s Frontier early‑access program on March 30, 2026. (microsoft.com) Microsoft describes Cowork as creating a plan, reasoning across Outlook, Teams, Excel and other work content, and carrying multi‑step tasks forward with visible progress and built‑in “skills” drawn from Anthropic and Microsoft toolchains. (microsoft.com) Researcher gained two named multi‑model capabilities—Critique and Council—that separate generation from evaluation and are implemented using combinations of models from Frontier partners, including Anthropic and OpenAI. (techcommunity.microsoft.com) Microsoft’s support docs say Critique runs a second reasoning pass (the “Critique model”) that applies Claude to strengthen structure, prioritize reputable sources, and add citation grounding, and that Critique becomes the default when Anthropic is enabled. (support.microsoft.com) Copilot Studio exposes a drop‑down model selector and multi‑agent tools so teams can mix Anthropic Sonnet 4 / Opus 4.1, OpenAI models, and entries from the Azure Model Catalog when building agents; Anthropic model choice began rolling out on Sept. 24, 2025 with admin opt‑in via Frontier. (microsoft.com) Microsoft announced a new Microsoft 365 E7 enterprise tier priced at $99 per user launching May 1, and a separate Agent 365 product priced at $15, with Microsoft saying E7 is roughly 65% more expensive than the current $60 E5 plan. (geekwire.com) Independent coverage and analysis flagged governance tradeoffs: Anthropic’s models are hosted outside Microsoft‑managed environments and commentators warned that AWS hosting can add latency, egress costs and data‑sovereignty complications that drive the need for model‑aware routing and monitoring. (microsoft.com) Microsoft and reporting outlets emphasize that Cowork runs inside Microsoft 365 security and compliance boundaries and that actions and outputs are auditable by default, and Microsoft cited Capital Group as an early access customer using Cowork in production scenarios. (microsoft.com)
