By Stuart Kerr, Technology Correspondent, LiveAIWire

Microsoft is now routing tens of thousands of AI prompts a week inside Excel and Outlook away from OpenAI and Anthropic and onto its own in-house models, according to Bloomberg, marking the first disclosed instance of the company shifting production-scale traffic off the two labs it has spent years and billions of dollars promoting. The change hasn’t been announced publicly. It was confirmed by a person familiar with the work, and Microsoft declined to comment when asked directly. But the timing says plenty on its own: this is happening just as the rest of Silicon Valley is going through its own AI spending hangover.

For most of the past year, big tech companies pushed employees to use as much AI as possible, some even running internal leaderboards to reward the heaviest users. That era, nicknamed “tokenmaxxing” inside the industry, appears to be ending fast. Microsoft’s quiet pivot toward its own MAI models isn’t an isolated cost-saving move. It’s the clearest evidence yet that the biggest AI spenders in the world are actively trying to spend less, not more, on the frontier models from OpenAI and Anthropic they were racing to adopt only months ago.

Why Microsoft Is Building Its Own Replacement for OpenAI and Anthropic

Microsoft has invested roughly $13 billion in OpenAI and, by its own account, already generates up to 30 percent of its code using generative AI tools. That scale of usage comes at a cost, and Mustafa Suleyman, the CEO of Microsoft AI, has been blunt about where he wants that cost to go. “We pay a lot of money to Anthropic,” Suleyman said in June, describing a goal to eventually eliminate that spending entirely. At Microsoft’s Build developer conference that same month, the company unveiled seven new MAI models spanning reasoning, coding, image generation, speech, and transcription, including one it says matches the coding performance of Anthropic’s Opus 4.6 model at a fraction of the price.

What’s notable is where Microsoft chose to start swapping models in. Excel and Outlook are two of the highest-volume, lowest-complexity workloads in the entire Microsoft 365 suite, exactly the kind of repetitive prompt traffic that racks up token costs without necessarily requiring frontier-model reasoning. Microsoft isn’t cutting OpenAI or Anthropic out of Copilot altogether. It’s demoting them from default infrastructure to a premium option reserved for harder tasks, while routing the routine, high-volume work to cheaper, in-house models it fully controls.

What This Means If Your Company Is Paying for AI Right Now

If Microsoft, with its $13 billion OpenAI stake, is actively working to reduce its own AI bill, that’s a signal worth paying attention to regardless of your company’s size. The broader pattern across the industry right now is a shift from unrestricted AI experimentation toward active cost governance: tracking which prompts actually produce value, and routing lower-stakes tasks to cheaper models rather than defaulting every request to the most expensive one available. If your organisation hasn’t yet audited which AI tasks genuinely need a frontier model versus a cheaper alternative, this is the moment several of the world’s largest tech companies are doing exactly that.

The Rest of Big Tech Is Having the Same Reckoning

Microsoft’s move lands in the middle of a much wider retreat from unrestrained AI spending. Meta’s chief technology officer, Andrew Bosworth, sent a memo to roughly 6,000 employees this summer after internal token usage saw what the company called an exponential increase in cost. “Nobody should be using AI tools just for the sake of using them,” Bosworth wrote, pushing back on the gamified leaderboards that had previously rewarded staff for burning through tokens on trivial tasks. Amazon and Meta have both since taken down internal usage leaderboards that once tracked who used the most AI.

Uber has been arguably the most exposed. The company burned through its entire 2026 AI coding budget in just four months and has since capped spending at $1,500 per employee per month per tool, even though nearly 70 percent of its committed code is now AI-generated. Consulting giant Accenture, meanwhile, has reportedly been trying to stop employees from using AI on basic tasks like converting PDFs into slide decks, after encouraging exactly that kind of heavy usage only months earlier. The reversal across all of these companies is the same: aggressive AI adoption without a clear way to measure whether it was actually paying for itself.

Why Some Companies Are Looking at Chinese AI Models Instead

The cost pressure has also pushed some companies toward cheaper alternatives out of China, including Z.ai’s GLM-5.2 and Moonshot AI’s Kimi, both of which are reported to cost a fraction of what comparable Western frontier models charge per token. That shift accelerated further after Anthropic’s Mythos 5 and Fable 5 models were briefly taken offline in June under a U.S. export control order, before access was restored on July 1. Coinbase CEO Brian Armstrong has said publicly that his company cut its AI usage costs in half after shifting some workloads to Chinese models.

That trend is not without friction. Anthropic has separately accused Alibaba of running a large-scale campaign to extract Claude’s capabilities through a technique called model distillation, alleging more than 28.8 million exchanges through roughly 25,000 fraudulent accounts between April and June. Alibaba denies the allegation, and none of Anthropic’s figures have been independently verified, so it’s worth treating this as an active, disputed claim rather than settled fact. It does, however, illustrate how directly the cost pressure driving companies toward cheaper models is now colliding with the security and intellectual-property concerns of the labs whose models are being undercut.

Where This Leaves OpenAI and Anthropic

None of this means OpenAI or Anthropic are going anywhere. Both remain central to Microsoft’s AI strategy for the hardest, highest-value tasks, and both reported record coding-tool revenue at the height of the tokenmaxxing boom. But the shift toward in-house and lower-cost alternatives, from Microsoft’s MAI models to cheaper Chinese options, suggests that growth engine may be entering a very different phase. The story going forward isn’t whether frontier AI labs get replaced. It’s whether they can keep growing once their biggest customers get serious about not overpaying for tasks a cheaper model could have handled just as well.

Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.