Microsoft has done something none of its competitors have publicly done: it shipped a production feature that uses OpenAI’s GPT and Anthropic’s Claude together in the same workflow, each doing what it does better than the other.

The feature is called Critique, and it lives inside Copilot Researcher, Microsoft’s deep-research agent for Microsoft 365 customers. The mechanics are straightforward: GPT drafts an initial research response, and Claude reviews it for accuracy, completeness, and citation quality before the result reaches the user. On the DRACO benchmark, which measures the quality of AI-generated research responses, this two-model architecture improved performance by 13.8 percent over any single-model approach.

That number outperforms standalone deep-research tools from OpenAI, Google, Perplexity, and Anthropic. The individual tools that supply each half of the Microsoft system each lose to the system they jointly power.

Why this matters beyond the benchmark

The result is a clear signal about where enterprise AI architecture is heading. Businesses and technology teams have spent the last two years debating which AI model to commit to. OpenAI versus Anthropic. Google versus everyone. The assumption embedded in that debate is that you pick one and deploy it.

Microsoft’s Critique feature challenges that assumption directly. If a draft-and-review sequence using two competing models produces meaningfully better output than either model alone, the practical question for every business deploying AI is not which model to use. It is which model for which step in which workflow.

This is not a new idea in software engineering. Composing specialist tools rather than relying on a single monolithic one has been best practice for decades. But in AI, the mythology of the universal model — the one that does everything — has been commercially powerful and intellectually persistent. The Microsoft result is one of the first production deployments to quantify what composability actually gains.

What this means for how enterprises should think about AI

For business leaders evaluating AI investments, the Critique feature has three direct implications.

First, vendor loyalty to a single AI provider is increasingly a strategic liability. If the best research output comes from combining GPT and Claude, and the best voice output comes from a different provider, and the best structured data work comes from a different model still, then enterprises that locked into a single-vendor strategy will find themselves working around a self-imposed constraint.

Second, the evaluation criteria for AI tools need to change. Benchmarking a model in isolation tells you what it can do. Benchmarking it in composition tells you what it enables. The DRACO improvement from draft-and-critique is not just a GPT improvement or a Claude improvement. It is an architectural improvement. Businesses evaluating AI vendors should be asking how their tools compose with others, not just how they perform alone.

Third, the operational skills required to get value from AI are shifting toward orchestration. Setting up a pipeline that routes tasks to the right model at the right step is more sophisticated than simply prompting a single AI assistant. Businesses that develop that orchestration capability — either internally or through partners — will have a meaningful edge over those still treating AI as a single-model tool.

The vendor strategy underneath the product decision

There is something else worth noting about this announcement that goes beyond the technical architecture.

Microsoft, which has invested over $13 billion in OpenAI, shipped a product that uses Anthropic’s Claude as a quality layer for OpenAI’s GPT output. That is not a vote of no confidence in OpenAI. But it is a public statement that Microsoft is willing to use whatever model produces the best outcome for customers, regardless of its investment relationships.

This is probably the right approach for a company serving enterprise customers at scale. What is interesting is that it contradicts the narrative that the AI industry is consolidating around a few dominant platforms. If the biggest enterprise AI platform in the world is orchestrating across competing models, the industry structure is more open than a winner-takes-all framing suggests.

For businesses making AI infrastructure decisions, that openness is good news. You do not have to guess which model wins. You can design systems that use the right model for each task and adjust as the landscape evolves.

What This Means for Business

The GPT-draft-Claude-critique architecture is not something most businesses will build themselves. But the principle behind it is immediately applicable.

Any business deploying AI agents should ask: is there a step in this workflow where a second model reviewing the first model’s output would catch errors or improve quality? For anything involving research, client communications, legal documents, compliance checks, or customer-facing content, the answer is almost certainly yes.

The cost of running a second model review on high-stakes outputs is trivial compared to the cost of a factual error in a client deliverable or a compliance failure in a regulated output. Microsoft’s result suggests the quality improvement is real. The operational question is how to structure that review within your existing workflows.

The practical next step is the free Working With Claude field guide. Thirty-two pages covering the ecosystem, Claude Code, and how to govern a rollout properly. Get your copy.