Microsoft puts OpenAI and Anthropic models on the same research desk

When you write an in-depth research report, your biggest enemy isn't a lack of information. It's confirmation bias. Language models have the exact same problem: when a single model plans, drafts, and reviews its own output, it reliably glosses over its own factual holes.

Microsoft rolled out a major update to Researcher, the deep research agent inside Microsoft 365 Copilot. The upgrade introduces two multi-model capabilities: Critique and Council. Instead of relying on a single engine to do the heavy lifting, Microsoft is now pairing models from OpenAI and Anthropic on the very same desk.

Splitting the drafter and the critic

The idea behind Critique is straightforward: separate generation from evaluation.

Traditional research agents rely on one model to handle research planning, web retrieval, synthesis, and writing. Critique divides that workflow between two distinct AI partners. One model leads the generative phase, searching sources and assembling an initial draft. A second model acts strictly as an expert reviewer, analyzing the draft before anything reaches the user.

This reviewer doesn't rewrite the piece from scratch. Instead, it grades the draft against structured rubrics: checking source reliability, confirming the report comprehensively answers the prompt, and enforcing strict evidence grounding so every major claim is tied to a verifiable citation. When users keep the model selector on Auto, Critique is now the default mode in Researcher.

To prove it actually works, Microsoft evaluated the setup on DRACO (Deep Research Accuracy, Completeness, and Objectivity), a benchmark released in February 2026 by researchers from Perplexity and academia across 100 complex real-world tasks. Using OpenAI's GPT-5.2 as the evaluation judge across five independent runs per question, Critique lifted Researcher's overall benchmark score by 13.8%.

The biggest jumps came in analytical breadth and depth (+3.33 points), followed by presentation quality (+3.04) and factual accuracy (+2.58), all with high statistical significance.

Side-by-side debate and Cowork

While Critique runs behind the scenes, Microsoft also introduced Council for users who want to see the friction directly.

Complex business topics rarely have one clean answer. Council displays responses from multiple frontier models side by side, accompanied by a synthesis cover letter. That letter spells out where the models agree, where their interpretations split, and what unique data points each model surfaced.

Alongside Researcher's updates, Microsoft also opened Copilot Cowork to its Frontier preview program. Built on the underlying platform behind Claude Cowork, it handles multi-step, long-running operational tasks like monthly budget reviews and schedule coordination. Capital Group reported using early access builds to prepare executive reviews and automate routine workflows.

Why it matters

For enterprise knowledge workers drowning in market analysis and executive briefing prep, the real hurdle with deep research tools has never been length. It has always been trust. Cross-checking citations from an AI report often eats up more time than writing the memo by hand. By separating the author from the editor, Critique introduces a sanity check that brings AI research much closer to production-ready corporate standards.

Strategically, this is Microsoft playing the role of the pragmatic enterprise aggregator. Despite billions invested into OpenAI, Microsoft is openly integrating models from Anthropic and other frontier labs directly into its core productivity suite. In our view, this signals a broader shift across enterprise software: the era of the monolithic single-model agent is fading, replaced by multi-model orchestration where different AI systems fact-check one another.

The obvious tradeoff here is latency and compute. Running multiple frontier models sequentially, combined with iterative critique loops, means research queries will take noticeably longer and burn significantly more tokens. For high-stakes decisions and market deep-dives, that trade is well worth making, but you won't want to spin up this engine for everyday quick questions.