Models are commoditizing faster than anyone expected. The companies that locked themselves into exclusive AI relationships are now paying premiums to escape them. And the clearest proof just landed inside the platform where 450 million people already work.
On March 9, 2026, Microsoft announced Copilot Cowork, a new agentic system that executes multi-step workflows across Outlook, Teams, Excel, and the rest of Microsoft 365. The technology powering it isn’t OpenAI’s. It’s Anthropic’s Claude.
This isn’t a story about Microsoft abandoning OpenAI. It’s about what happens when a company that has poured billions into a single AI provider decides exclusivity is a liability—not an asset. And if you’re an IT leader planning your next Microsoft licensing move, the implications are immediate.
What Actually Happened
The Technical Move
Copilot Cowork is fundamentally different from earlier Copilot versions. Previous iterations drafted emails, summarized documents, answered questions. Cowork executes. Users describe an outcome. The system builds a plan, pulls context from across Microsoft 365 (emails, meetings, files, calendar) and runs the job in the background with transparent checkpoints.
From the demo:
- Review a month of one-on-ones with direct reports, surface patterns, produce a briefing
- Pull emails, files, research, and calendar prep for a client meeting
- Build competitive analysis with supporting Word and Excel docs
- Triage and restructure an overbooked calendar
This isn’t asking Claude for a draft. It’s Claude doing the work autonomously, checking in at decision points, and delivering finished output.
The Deal Architecture
The timeline matters because it reveals just how deliberate this was.
November 2025: Microsoft and Nvidia co-invest in Anthropic, $5B and $10B respectively. Anthropic’s valuation jumps from $183B to $350B. Simultaneously, Anthropic commits to purchasing $30 billion in Azure compute, one of the largest infrastructure deals ever. Claude will now run on Azure, powered by Nvidia hardware (Grace CPUs, Blackwell/Vera Rubin GPUs).
Late 2025 onward: Claude gets integrated into Microsoft Foundry. Becomes an option in Copilot Chat for Frontier users. Microsoft deploys Claude Code internally across its own engineering teams.
March 2026: Copilot Cowork launches to the Frontier program. Claude is now the model for agentic, long-running tasks inside Microsoft 365.
On top of all this, Microsoft is spending approximately $500 million per year on Anthropic’s AI models, according to The Information. That level of commitment tells you this isn’t an experiment.
Why Claude, Not GPT?
The answer cuts to the heart of why this matters.
In benchmark testing, the multi-model approach (GPT drafting, Claude reviewing) scored 13.8% higher on DRACO than any single-model configuration, including standalone GPT and Claude. For complex reasoning and long-running agentic tasks, that gap is what drove the architectural decision.
Microsoft selected Claude specifically for the agentic tier, citing stronger performance on multi-step reasoning and long-running workflows. Spataro described Anthropic’s agentic harness as what “allows this new Copilot to actually manipulate software and perform complex workflows,” distinguishing it from what GPT had previously enabled.
For enterprise customers in regulated industries like finance, healthcare, and legal, this isn’t theoretical. When you’re running agents for hours across sensitive data, consistent reasoning matters more than raw speed.
Jared Spataro, Microsoft’s CMO for AI at Work, told Fortune:
“Every 60 days at least, there’s a new king of the hill. There’s so much demand for a platform that doesn’t feel like, ‘I have to skip over to the next vendor.'”
Microsoft is done betting the farm on any single model. And if the interest I’ve been seeing across LinkedIn and Twitter is any indication (Claude has become the most-discussed model among enterprise IT professionals over the past six months), customers are ready for that shift.
The OpenAI Relationship Has Fundamentally Changed
The Microsoft-OpenAI partnership was built on exclusivity and mutual dependency. That’s no longer true.
What Broke It
Starting in mid-2025, friction became visible. OpenAI wanted looser Microsoft control over its products. OpenAI needed Microsoft approval for its for-profit conversion. It was granted, but tensions ran high enough that OpenAI executives reportedly discussed accusing Microsoft of anticompetitive behavior as a negotiating tactic.
Meanwhile, OpenAI was aggressively diversifying its infrastructure away from Azure. Across three deals in 2025 (March, May, and September), OpenAI signed commitments with CoreWeave totaling $22.4 billion. It also pursued Google for server capacity.
October 2025 brought the restructuring deal: Microsoft received a 27% stake in OpenAI (valued at roughly $135B), OpenAI converted to for-profit with governance protections, and Microsoft retained access to OpenAI’s technology through 2032. AGI decisions became subject to independent review.
But friction didn’t stop. By March 2026, reports indicate OpenAI is building an internal code-hosting platform it may sell commercially, putting it in direct competition with GitHub, which Microsoft bought for $7.5 billion in 2018.
The New Reality
The partnership is now transactional, not exclusive. OpenAI committed to $250 billion in Azure compute purchases. Microsoft retains technology access through 2032 and holds a significant stake in OpenAI’s upside. But both are competing in enterprise productivity software, developer tools, and agentic AI. Neither fully controls the other.
Microsoft holds a 27% stake in a company that’s also becoming a competitor. This is what a bilateral relationship looks like when both sides outgrow the original arrangement.
The Adoption Problem That Made This Necessary
Here’s the context that makes Microsoft’s move strategic rather than opportunistic: Copilot hasn’t been a runaway success.
Despite massive infrastructure investment, Copilot represents roughly 3.3% of Microsoft’s commercial user base, about 15 million paid licenses across more than 450 million commercial subscribers. For IT leaders who’ve been fielding budget questions about Copilot ROI, that number isn’t surprising. The value proposition for general-purpose AI assistance has been a tough sell at $30/user/month on top of existing licensing.
Then Anthropic released Claude Cowork on January 13, 2026.
Enterprise buyers saw what a genuinely capable agentic system looked like. The gap between that and what Copilot had been offering became visible. A roughly $285 billion selloff in enterprise software stocks coincided with Claude Cowork’s launch and Anthropic’s subsequent release of open-source plugins in early February, as investors began pricing in the risk that AI agents could replace traditional SaaS workflows entirely.
Microsoft’s stock fell approximately 10% the day after its Q2 earnings report in January, driven by investor concern over slowing Azure growth and surging capital expenditure. The stock has shed roughly 14% since Anthropic launched Claude Cowork in mid-January, a period that captured both the earnings reaction and broader competitive anxiety about AI disrupting SaaS.
Microsoft’s response was pragmatic: absorb the threat rather than fight it. License the same technology, wrap it in enterprise infrastructure, embed it in the platform where the work already happens. Instead of losing to Claude Cowork, Microsoft became a Claude distribution partner.
What This Actually Means
For IT Decision-Makers
This is where it gets practical.
Microsoft 365 E7 launches May 1, 2026 at $99/user/month, bundling:
- Microsoft 365 E5 (core productivity suite)
- Microsoft 365 Copilot (general-purpose AI)
- Entra Suite (identity management)
- Agent 365 (AI governance and monitoring)
Bought separately, that’s $117. The E7 bundle saves $18/user/month, which is meaningful at scale.
The real value isn’t the discount, though. It’s the operational simplification. One platform, best-available models: Claude for complex reasoning and agentic tasks, GPT for general-purpose work, Microsoft’s own models where relevant. No separate vendor evaluations. No additional API keys. No parallel compliance reviews.
If you’re currently negotiating an E5 renewal or evaluating Copilot ROI, the E7 launch date should be on your calendar. The bundling economics change the conversation significantly.
And if you’re a CISO, Agent 365 deserves close attention. When AI agents can execute multi-step workflows across your organization’s email, calendar, and files, the governance question (what can these agents access, what can they do, and who approved it) becomes the real adoption barrier. Agent 365 is Microsoft’s answer, and how well it works will determine whether enterprises actually turn Cowork on.
The risk to this thesis is real, though: multi-model orchestration adds complexity, cost, and potential inconsistency. Microsoft’s own Council feature costs roughly 2.5 times a single-model query. If users experience slower, less coherent responses, or if IT teams struggle to govern which model is running what—the multi-model strategy could fragment the experience rather than enhance it. The bet works only if the platform layer is strong enough to make model-swapping invisible to end users.
For Anthropic
This is the single most significant enterprise distribution deal Anthropic has closed.
Anthropic’s trajectory tells the story: annualized revenue hit roughly $14 billion by its Series G in February 2026, up from a $9 billion run rate at end of 2025, and was reportedly approaching $19 billion by March. The company raised $30 billion in its Series G at a $380 billion valuation. Enterprise customers now account for roughly 80% of revenue.
Being embedded as the agentic model inside Microsoft 365, used by hundreds of millions of workers and 90% of the Fortune 500, is distribution that Anthropic’s sales team couldn’t build alone. Claude now runs on every major cloud: AWS Bedrock, Google Vertex AI, and Microsoft Azure. Enterprise customers have maximum flexibility regardless of their existing cloud commitments.
For OpenAI
OpenAI isn’t losing Microsoft. It’s losing exclusivity. That distinction matters, but so does the trajectory.
OpenAI’s position remains formidable: annual recurring revenue above $20 billion, a funding round that closed March 31 at $122 billion (the largest private investment in history), and a post-money valuation of approximately $852 billion. The Azure commitments and the ongoing integration of GPT models into Copilot remain intact.
But when Jared Spataro says “we’re model-agnostic,” he’s sending a message directed at OpenAI as much as at customers. No single model will dominate Copilot. For a company that spent six years positioning itself as Microsoft’s AI foundation, that’s a significant shift.
The Bigger Pattern
This isn’t about Claude vs. GPT. It’s about what happens when a market matures past single-vendor dependency, and what that means for how enterprises should think about AI strategy.
The enterprise AI platform of the future doesn’t lock you into a single model provider. The durable value sits in the platform where work happens, the data integration connecting your systems, the security perimeter governing access, and the workflow context that understands what work actually matters.
Models are trending toward interchangeability at the platform level—not because output quality no longer varies, but because the platform is increasingly responsible for routing the right model to the right task. The model you see is becoming an implementation detail. The platform is the product.
For IT leaders, the implication is straightforward: stop evaluating AI models in isolation. Evaluate the platforms that orchestrate them. The model layer will keep churning every 60 days, just as Spataro said. The platform, the governance, the data integration—that’s where your investment compounds.
What Happens Next
Copilot Cowork is in Frontier (early access) now. Broader availability is expected later in Q2 2026.
Microsoft 365 E7 launches May 1, 2026 at $99/user/month.
Agent 365 will be the sleeper story. The trust question—”can I let an AI agent actually execute work across my organization?”—is what separates pilot programs from production deployments. How Microsoft answers it will determine whether Copilot Cowork becomes the platform’s killer feature or another expensive capability that IT keeps turned off.
The next twelve months will show whether Microsoft can deliver on the promise of best-available models inside the platform you already pay for, at scale. If they do, every enterprise software vendor starts rethinking their AI strategy. If they don’t, Copilot Cowork becomes another adoption problem, and Anthropic keeps building distribution as a standalone platform.
The Bottom Line:
Microsoft just made AI exclusivity obsolete. For enterprise buyers, choice is now the baseline, and the licensing economics finally reflect it.
Whether this is good for OpenAI, Anthropic, or Microsoft over the next five years depends entirely on execution. But for IT leaders making platform decisions today, the calculus just changed. Evaluate accordingly.
I’d like to hear what you’re seeing in the field. Is your organization evaluating Claude alongside GPT? Has the E7 bundling changed your licensing conversations? Drop a comment or reach out. These are the discussions that shape how we all navigate what’s coming next.
— Bharat