Microsoft adds 'Critique' to Copilot
Microsoft launched 'Critique' inside Copilot, leveraging multiple model backends (reportedly including Anthropic and OpenAI) to boost research accuracy by about seven points in early demos — a practical nudge toward hybrid multi‑model production systems. The release highlights an industry trend toward ensemble routing and model specialization in shipped products. (x.com) (x.com)
Microsoft published the Copilot Cowork announcement and the Researcher multi‑model update on March 30, 2026, and said Cowork is available via its Frontier early‑access program. (microsoft.com) Microsoft’s product post and Community Hub entry describe Critique as a two‑stage workflow that separates generation from evaluation, assigning drafting and iterative retrieval to one model and a distinct review/refinement role to a second model. (techcommunity.microsoft.com) The company frames its comparison results against the newly released DRACO benchmark, a 100‑task, 10‑domain suite authored by Perplexity and Harvard collaborators, and notes that all baseline scores come from the DRACO paper except the Researcher‑with‑Critique evaluation. (arxiv.org) Microsoft says Critique will be the default experience when “Auto” is selected in the Researcher model picker, and the companion Council mode surfaces side‑by‑side model reports with an automated cover letter that highlights where models agree and diverge. (techcommunity.microsoft.com) The Copilot Cowork write‑up stresses a “multi‑model advantage” and the reuse of technologies from Claude Cowork to handle long‑running, multi‑step workflows, with customer pilots (including Capital Group) cited as early adopters. (microsoft.com) Industry coverage framed the rollout as a concrete move toward model‑agnostic orchestration and ensemble routing in shipped products, characterizing the Researcher updates as part of a broader vendor trend to specialize roles across models rather than rely on single‑model stacks. (thenewstack.io)
