$13 billion. That’s how much Microsoft invested in OpenAI, yet here they are, hedging their bets by bringing Anthropic’s Claude into Copilot. If that doesn’t tell you everything about the current state of AI model reliability, I don’t know what will.
Microsoft’s latest Copilot update does something I never thought I’d see: it pits OpenAI’s GPT against Anthropic’s Claude in a corporate cage match. The new “Critique” feature has GPT draft responses while Claude fact-checks them. It’s like hiring two contractors because you don’t trust either one to finish the job alone.
Why This Actually Matters
Look, I’ve tested dozens of AI tools, and they all have the same dirty secret: they hallucinate. They make stuff up. They sound confident while being completely wrong. Microsoft knows this, which is why they’re not putting all their eggs in the OpenAI basket anymore.
The Copilot Researcher now lets users choose between GPT and Claude for research tasks. There’s also a “Council” feature where both models weigh in on complex queries. Microsoft is essentially admitting that no single AI model is good enough on its own.
And honestly? That’s the most honest thing a major tech company has said about AI in months.
The Real Strategy Here
Microsoft isn’t playing favorites anymore. They’re betting that their advantage isn’t in having the best model, but in having the best data and integration. They’ve got access to your emails, documents, calendars, and entire workflow through Microsoft 365. That’s the moat, not the model.
This explains why they launched Copilot Cowork, an enterprise AI agent built entirely on Anthropic’s technology. They’re not married to OpenAI anymore. They’re model-agnostic, which is probably how every company should be thinking about AI right now.
What This Means for Users
If you’re using Copilot for research, you now have options. Want GPT’s creative flair? Use that. Need Claude’s more measured, analytical approach? Switch over. Want both to argue it out? Let them duke it out in Council mode.
The Critique feature is particularly interesting. GPT generates the initial response, then Claude reviews it for accuracy and flags potential issues. It’s like having a second pair of eyes, except both pairs of eyes are artificial and occasionally unreliable.
But two unreliable AI models checking each other? That might actually add up to something more dependable. Maybe.
The Uncomfortable Truth
Microsoft’s move exposes something the AI hype machine doesn’t want you to know: these models aren’t as good as the marketing suggests. If they were, Microsoft wouldn’t need to run them in parallel and have them fact-check each other.
This isn’t a feature. It’s a workaround.
Every AI tool I review on agnthq.com gets the same treatment: I test it until it breaks. And they all break. GPT hallucinates sources. Claude gets overly cautious and refuses to answer straightforward questions. They each have different failure modes.
Microsoft is essentially productizing the workaround that power users have been doing manually: asking the same question to multiple AI models and comparing answers.
Should You Care?
If you’re already paying for Microsoft 365 Copilot (and it’s not cheap), this upgrade gives you more tools to work with. The ability to choose your model or have multiple models collaborate could genuinely improve your results.
But if you’re deciding whether to buy into Copilot based on this news, pump the brakes. This update is Microsoft admitting they need multiple AI providers to deliver reliable results. That’s not exactly a ringing endorsement of AI’s current capabilities.
The smart play here is to watch how this performs in the real world. Microsoft is doing the expensive experiment of running multiple frontier models simultaneously. Let them figure out if it actually works before you commit your company’s budget to it.
For now, Microsoft’s multi-model approach is the most pragmatic thing happening in enterprise AI. It’s not sexy, it’s not simple, but it might actually be useful. And in the world of AI tools, “actually useful” is a higher bar than you’d think.
🕒 Published: