The race to build the most capable AI research assistant has intensified throughout the past year. Google launched its research agent for Gemini in December 2024, followed by OpenAI in February 2025, then xAI and Perplexity shortly after. Anthropic‘s Claude also introduced its own agent in April of last year, building a following among professionals who rely on detailed, cited outputs. Each company has positioned its individual model as the superior research tool.
Microsoft is taking a different approach entirely. On Monday, the company announced two new features for its Copilot Researcher tool — called Critique and Council — that deploy both OpenAI’s GPT and Anthropic’s Claude on the same research task rather than relying on a single model. According to Microsoft’s own testing against an industry benchmark, the combined system outperforms every other model included in that evaluation.
Critique addresses a fundamental limitation shared by all current AI research tools: a single model plans, searches, drafts, and delivers a report with no independent review of its own work. This workflow leaves room for hallucinations, citation errors, and inaccurate claims to reach the user unchecked. Critique splits the process into two distinct phases to reduce these risks. GPT handles the first stage — planning the research, retrieving sources, and producing an initial draft — while Claude then acts as a strict editor, assessing factual accuracy, citation quality, and whether the response fully addresses the original question.
Microsoft says the roles can eventually be reversed, with Claude drafting and GPT critiquing, though GPT currently leads the generation phase. Satya Nadella, Microsoft’s chief executive, described Critique on social media as “a new multi-model deep research system” that uses “a combination of models from Frontier labs, including Anthropic and OpenAI.” The company explains that separating generation from evaluation is central to the design, with one model producing the initial work and a second functioning as an expert reviewer before the final report is delivered.
On the DRACO benchmark — a standardized evaluation covering 100 complex research tasks across 10 domains including medicine, law, and technology — Copilot with Critique scored 57.4 points. Claude Opus 4.6 operating alone reached 42.7 points, meaning Microsoft’s combined system exceeds the next best result by nearly 14 percentage points. The largest improvements appeared in breadth of analysis and presentation quality, with factual accuracy also showing a notable gain.
The second new feature, Council, takes a parallel rather than sequential approach. Instead of one model reviewing the other’s draft, Council runs GPT and Claude at the same time and places their complete reports side by side. A third model then reads both outputs and produces a summary identifying where the two systems agreed, where they diverged, and what distinct angles each one surfaced that the other did not. This automates a comparison process that users have previously had to carry out manually.
The two features operate differently within Copilot’s interface. Critique is the default experience inside Researcher, while Council must be activated by selecting “Model Council” from a picker. Both are currently available to users enrolled in Microsoft’s Frontier program, the early-access channel for Copilot’s newest capabilities. Access requires a Microsoft 365 Copilot license at $30 per user per month, along with enrollment in the Frontier program. Microsoft’s broader strategic position, despite its multibillion-dollar partnership with OpenAI, appears to rest on the view that no single model will hold the top position indefinitely, and that the greater long-term value lies in an orchestration layer capable of routing tasks to whichever model combination performs best.
Originally reported by Decrypt.
