Microsoft’s “alliance of the century” with OpenAI is being quietly rewritten. According to multiple reports, Microsoft has begun routing tens of thousands of AI prompts in core products like Excel and Outlook away from OpenAI’s models to its in-house MAI (Microsoft AI) models to slash inference costs. Almost simultaneously, Microsoft’s first homegrown cybersecurity AI model, MAI-Cyber-1-Flash, beat Anthropic’s and Google’s flagships on the CyberGym vulnerability benchmark — at half the cost. The era of “coopetition” between the two giants has officially entered its second half.

Tens of Thousands of Prompts Rerouted: Up to 89% Cost Savings

According to The Information and follow-up coverage, Microsoft CEO Satya Nadella laid out the company’s reasoning in a lengthy X post: why build your own AI models instead of leaning on OpenAI and Anthropic? The logic is blunt: cost and control. A large share of routine AI tasks inside Excel and Outlook are now handled by MAI-series models — tasks that don’t need frontier-level reasoning yet consume enormous inference budgets. VentureBeat cites Microsoft data showing the new MAI-Flash models, designed for high-efficiency, low-cost workloads, cut GPU costs by up to 89% versus calling OpenAI models directly, and already power Bing, Excel, Copilot, and Dynamics 365.

The strategy is notably nuanced: Microsoft is not cutting over from OpenAI wholesale. As Memeburn reports, GPT-5.6 remains the “preferred model” powering Microsoft 365 Copilot across Word, Excel, and PowerPoint — a status OpenAI celebrated alongside the model’s public launch on July 9. What Microsoft is building instead is a multi-model routing architecture: workloads are dynamically assigned by complexity, latency requirements, and cost-efficiency. Simple tasks go to its own lightweight models; only the hardest frontier problems are routed to OpenAI’s or Anthropic’s flagship systems. Forbes describes it as bringing “most security scanning” in-house while outsourcing only the hardest work.

Beating Anthropic and Google at Half Price: What Is MAI-Cyber-1-Flash?

Meanwhile, Microsoft has released its first in-house cybersecurity AI model, MAI-Cyber-1-Flash, entering public preview alongside the agentic defense platform code-named Project Perception. According to Tech Times and CNET, the model scores 95.95% on the CyberGym vulnerability benchmark, claiming to outperform Anthropic’s Claude lineup and Google’s Gemini. CNBC notes that when paired with OpenAI’s general-purpose GPT-5.4, it performs even better — at roughly half the cost of competing solutions.

Microsoft’s confidence stems from its vast security-data estate. As Ars Technica reports, the company processes telemetry at a scale no single rival can match — more than 80 trillion signals daily across endpoints, identities, email, and cloud. Project Perception is built around “agentic defense”: AI agents continuously scan codebases and digital environments to proactively identify, assess, and remediate risks, rather than waiting for human instructions. VentureBeat notes the platform combines low-cost models with Microsoft’s proprietary security data to push the unit economics of enterprise security operations dramatically lower.

Nadella’s Warning: AI Monopoly Could “Hollow Out” Industries

Beneath the deployments, Nadella’s public rhetoric has grown sharper. As Memeburn reports, he warned that AI monopoly risk is real: if just a few giant models capture all the value, entire industries could be “hollowed out” — much as globalization once gutted manufacturing. The remark is striking coming from Microsoft’s helm: the company is simultaneously OpenAI’s largest investor and its biggest enterprise customer, yet Nadella clearly has no intention of betting Microsoft’s AI lifeline on a single partner.

At the same time, Microsoft is diversifying its partner portfolio. According to The Information, Microsoft will pay to use Anthropic’s technology for some AI features in Office 365 apps — another clear signal of the software giant’s supply-chain diversification.

What It Means: Goodbye, Single-Model Era

Microsoft’s moves reveal a bigger industry shift: the future of AI applications belongs not to a single model, but to the routing layer. As model capabilities converge, the decisive advantage shifts to who can deliver the right task to the right model at the lowest cost and lowest latency. Microsoft owns the world’s largest enterprise distribution network — Windows, Office, Azure — and once its in-house models reach “good enough,” economies of scale automatically amplify the cost advantage.

For OpenAI, it is a dangerous signal: its biggest customer is becoming its strongest competitor. For developers and enterprise users, it may be good news — fiercer competition usually means lower prices and more choices. For investors, the real question is: as the “frontier premium” gets eroded by cheap in-house models, how long can OpenAI’s hundreds-of-billions valuation narrative hold?

Conclusion

Microsoft hasn’t announced a “breakup” with OpenAI, but actions speak louder than words: what it can build itself, it builds; where cheaper options exist, it uses them. Each rerouted AI prompt inside Excel accumulates into a silent transfer of power. For enterprise decision-makers watching their AI vendor strategy, Microsoft’s textbook answer is: don’t bet on a single model — build your own routing layer. This cost war among giants has only just begun.