Microsoft Copilot Cowork launches, with new features letting the GPT and Claude dual models review each other

Microsoft announced a major update to Microsoft 365 Copilot on Monday, launching two breakthrough features at once: a multi-model research architecture that integrates GPT and Claude, and the intelligent agent Copilot Cowork, which can autonomously track long-running tasks. Both features are being opened to select users through the Frontier early access program.

Deep research AI competition is heating up—Microsoft rolls out a cross-model strategy

Over the past year, deep research has become one of the most fiercely contested battlegrounds in the AI field. Google was first to introduce research agent functionality for Gemini. OpenAI followed with its own version. xAI and Perplexity have also moved in succession, and Anthropic’s Claude has built a strong track record among professional user groups.

Every company is working to convince users that its own single model is the smartest AI research assistant. But Microsoft has gone in the opposite direction, asking, “Why choose only one model?”

Nearly all AI research tools on the market follow the same pattern: “user prompts → a single model plans the search → crawls and organizes sources → writes the report.” Because the entire process is handled end-to-end by one model, it can lead to hallucinations, incorrect citations, and false or inaccurate statements.

Microsoft’s two newly launched research features are designed as solutions specifically targeting this weakness.

Critique: a review mechanism where GPT generates and Claude verifies

The first feature is called Critique. Its core idea is to split the research workflow into two parts—generation and evaluation—handled by different models, introducing a mechanism similar to “peer review.”

More specifically, GPT is responsible for outlining research directions, searching for sources, consolidating references, and drafting the report. After that, Claude reviews the draft one by one across multiple dimensions—source credibility, report completeness, factual support, and more—strengthening report quality before delivering the revised version to the user. Microsoft says it also plans to support a reversed configuration in the future, with Claude generating and GPT reviewing.

In testing performance, Critique has been highly impressive. For example, using the DRACO benchmark test that covers ten major domains—medicine, law, technology, and others—and 100 complex research tasks in total, Copilot Researcher with Critique scored 57.4, while Claude Opus alone scored 42.7, with a lead of nearly 14%.

Council: GPT and Claude side-by-side, with a judge model evaluating

The second feature, Council, takes a completely different approach. Once enabled, GPT and Claude will both complete a complete research report at the same time, independently. The two reports are presented side by side to users, making it easy to see differences in factual citations, analytical angles, and interpretations of the data between the models.

After that, a third “judge model” reads both reports, compiles a consolidated summary, and explains where the two agree, where they differ, and how they can complement each other.

Copilot Cowork: you give the task, and the AI agent runs it on its own

At the same time, Microsoft also announced the long-awaited Copilot Cowork feature rollout. As an intelligent agent that can operate autonomously within the Microsoft 365 ecosystem, Cowork is designed specifically for enterprise workflows that require “long-running, multi-step” execution.

Users only need to specify the task objective. Cowork will proactively track progress, plan the next actions, and keep pushing the work forward without requiring the user to stay on and monitor the entire process. For complex projects that may take hours or even days to complete, Cowork lets users rest easy with “assign it and let it run,” making it a trustworthy work partner.

Who can use Copilot Cowork, Critique, and Council?

Critique, Council, and Copilot Cowork are currently all available through Microsoft’s Frontier early access program. Frontier is Microsoft’s early-access channel aimed at enterprise customers. Users must pay for a Microsoft 365 Copilot subscription and also join the Frontier program.

Critique automatically enables in the “Auto” mode. Council, on the other hand, requires users to manually switch to “Model Council” in the model selector.

Microsoft is betting not on a model, but on AI orchestration and autonomy

Even though Microsoft and OpenAI have a long-term partnership, Microsoft still chooses to bring in Claude—reflecting a deeper strategic judgment: no single AI model can stay in the lead forever.

For companies evaluating enterprise AI tools, this time Microsoft’s rollout of these capabilities may be redefining what an AI work partner is: the strongest AI colleague has never been a single model, but rather a system that lets the best models handle their respective roles—and keeps working even when you’re away.

This article, Microsoft Copilot Cowork goes live—new features add mutual review between GPT and Claude, first appeared on Chain News ABMedia.