Skip to content

AI Cost Optimization

Microsoft just started swapping the AI behind Copilot for its own, cheaper models. Here is what business owners should know.

On July 7, 2026, Bloomberg reported that Microsoft has started replacing OpenAI and Anthropic models with its own internally built MAI models inside Excel and Outlook Copilot, tens of thousands of prompts a week so far, still a small share of total usage. Microsoft's AI chief, Mustafa Suleyman, confirmed the reason plainly: 'We pay a lot of money to Anthropic, so our goal is to reduce and ultimately eliminate that cost.' The direct answer for a business owner: the Copilot button in your Excel and Outlook looks the same as it did last month, but the model answering behind it may not be, and Microsoft is not required to tell you when that changes. If your business runs any AI feature bundled inside software you already pay for, this is a live example of what that arrangement can look like once the vendor decides to protect its own margin.

By Fabio Rabelo · Founder, ATLACIS ·

What happened

Bloomberg reported that Microsoft has started routing a meaningful and growing number of Copilot tasks inside Excel and Outlook to its own MAI models rather than the OpenAI and Anthropic models those apps previously relied on more heavily. According to a person familiar with the work, tens of thousands of AI prompts a week in the two applications are now completed by Microsoft's internal models. Microsoft itself did not confirm the details, and a spokesperson declined to comment on the report. The reason behind the shift is not disputed. Mustafa Suleyman, Microsoft's AI chief, said in remarks confirmed by Bloomberg and separately by SiliconANGLE: 'We pay a lot of money to Anthropic, so our goal is to reduce and ultimately eliminate that cost,' adding in a separate interview that 'Anthropic is extremely expensive and I think many people are urgently looking for alternatives.' This follows Microsoft's June Build developer conference, where it introduced seven new MAI models, including a reasoning model called MAI-Thinking-1 and a coding model Microsoft said could match the coding performance of Anthropic's Opus 4.6 at a lower cost. MAI models are also already available inside GitHub Copilot, and Microsoft has said a company-built transcription model is coming to Microsoft Teams.

Why it matters for business owners

Most businesses running Microsoft 365 Copilot did not choose OpenAI or Anthropic specifically. They bought Copilot, and Copilot happened to run on those models behind the scenes. That same arrangement now means Microsoft can substitute its own models for the same feature, at the same price to the customer, for reasons that have nothing to do with whether the new model is better suited to a given task. The Decoder's reporting adds an important detail: independent benchmarks published around Microsoft's June announcement showed MAI-Thinking-1 trailing OpenAI's and Anthropic's newest models, performing roughly on par with DeepSeek's V3.2, a more modest result than Microsoft's own comparison against an older Anthropic model. Whether or not that gap closes over time, the immediate point for an owner is this: a bundled AI feature is not a fixed product. The vendor can and will change what powers it, on its own schedule, driven by its own cost pressure, and a business paying the same subscription fee has no built-in way to know when that happened.

What owners should not misunderstand

This is not a story about Microsoft doing something improper. Multi-model platforms are common across the industry, and choosing the most cost-effective model for a given request, a practice generally called model routing, is a legitimate and increasingly standard way vendors manage compute cost at scale. Microsoft has not confirmed exactly which features moved, how its 'Auto' or default mode is weighted, or what share of total Copilot traffic MAI models now handle. Reporting is explicit that this remains a small fraction of Microsoft's overall AI usage today, and Microsoft continues to license OpenAI and Anthropic models broadly, including a discounted OpenAI arrangement that reportedly runs through 2032. Do not read this as Microsoft abandoning its frontier-lab partners, and do not assume every Copilot response you get today is coming from a weaker model. The accurate read is narrower and, for planning purposes, more useful: the model behind a bundled AI feature can change without a corresponding change in your bill, your contract, or the interface you see.

The operational lesson

When an AI feature is bundled into software you already pay for, 'AI included' is not one fixed product. It is a moving target the vendor controls, and the vendor's incentive to route your requests to the cheapest available model is not the same as your incentive to get the best available answer for a specific task. This matters most where the specific model behind a result actually affects the outcome: technical or coding work where capability gaps show up in edge cases, any output tied to a compliance or data-handling claim about a specific model, or a workflow your team has already tuned prompts around a particular model's behavior. It matters less for routine, low-stakes drafting where any competent model produces an acceptable result. The mistake is assuming a bundled AI feature is a stable, single-vendor product just because the interface never changes. It is closer to a utility with a variable, vendor-controlled supply behind the meter.

What a serious business should do next

Ask your primary AI-bundled software vendors, not just Microsoft, whether their platform uses multi-model routing on the tier you are paying for, and whether you can see or control which model answered a given request. For any workflow where the specific model genuinely matters, check whether your subscription tier lets you pin a named model instead of relying on 'Auto' or default settings, and use that option where the stakes justify it. Periodically retest a small, consistent set of real tasks through your bundled AI features and compare the output quality over time, the same way you would monitor any vendor's service level rather than assuming a flat subscription means a flat product. Do not treat this as a reason to disable or distrust Copilot broadly. Reporting is clear that this change today affects a small share of total usage, and MAI models still need to prove themselves against the field over time. Treat it instead as a prompt to add one recurring question to your AI vendor reviews: not just what does this cost, but what is actually answering, and has that changed since I last checked.

The Atlacis view

Atlacis does not take a position on whether Microsoft's own models are as good as OpenAI's or Anthropic's, and that comparison will keep shifting regardless. What matters to a business owner is a pattern Atlacis sees repeatedly: AI capability arrives bundled inside software a business already trusts, and the specific model doing the work is treated as an implementation detail rather than a decision the owner ever consciously made. That was tolerable while the bundled feature was a novelty. It stops being tolerable once a workflow depends on consistent output, or once a claim about data handling or model behavior is tied to a specific model that quietly stopped being the one running the task. Atlacis helps owners map which models actually sit behind the AI tools they already use, decide where that choice needs to be pinned down rather than left to the vendor's default, and build that question into how they evaluate any AI-bundled software going forward.

The short version

  • Bloomberg reported on July 7, 2026 that Microsoft has started routing some Excel and Outlook Copilot tasks to its own MAI models instead of OpenAI or Anthropic, still a small share of total usage but a stated cost-reduction strategy.
  • Microsoft AI chief Mustafa Suleyman confirmed the motive directly: reducing and ultimately eliminating what Microsoft pays Anthropic, which he separately called 'extremely expensive.'
  • Independent benchmark reporting around Microsoft's June Build conference showed its new MAI-Thinking-1 model trailing OpenAI's and Anthropic's latest models, roughly on par with DeepSeek V3.2, a more modest result than Microsoft's own comparison against an older Anthropic model.
  • A bundled AI feature inside software you already pay for is not a fixed product. The vendor can change the model behind it, for its own cost reasons, without changing your price, your contract, or the interface you see.
  • For workflows where the specific model matters, check whether your subscription tier lets you pin a named model instead of a default or Auto setting, and periodically retest output quality rather than assuming a flat subscription means a flat product.
Tags:AI costvendor dependencyMicrosoftCopilotmodel routingAI vendorsbusiness AIAI decision-makingAI buying decisionsvendor lock-in
FAQ

Common questions

Did Microsoft stop using OpenAI and Anthropic models in Copilot?
No. Reporting confirms this change affects only a small share of Microsoft's overall Copilot traffic so far, concentrated in some Excel and Outlook tasks. Microsoft continues to license OpenAI and Anthropic models broadly, including a discounted OpenAI pricing arrangement reported to run through 2032. The direction of travel toward more in-house model usage is explicit, but this is not a full replacement today.
Should my business stop using Microsoft Copilot because of this?
Not automatically. The practical response is not to switch tools, it is to add a question to how you evaluate any AI feature bundled into software you already pay for: which model is actually answering, can you see or control that, and does it matter for the specific workflow. For low-stakes drafting, this change is unlikely to be noticeable. For workflows where model-specific behavior matters, check whether your subscription tier allows pinning a named model instead of a default setting.
Keep reading

More from the blog

Microsoft just changed how it bills for office AI. Here is what it means for your budget.

Microsoft launched Copilot Cowork worldwide on June 16, 2026, an AI agent that completes complex office tasks autonomously. For the first time in roughly two decades, Microsoft changed its pricing model for this capability: usage is now billed by task through a system called Copilot Credits. For business owners running Microsoft 365, the cost model for AI just changed, and it needs to be understood before the feature is enabled.

Uber burned through its annual AI budget in four months. Here is what business owners should know about AI cost controls.

CNBC reported this week that enterprise AI spending is hitting a wall. Uber burned through its annual AI budget in four months before capping employee access at $1,500 a month. A 25-person startup switched its entire AI stack off Claude because AI costs had exceeded its payroll. The era of unchecked AI spending is ending at the top. For most business owners, the lesson is simpler: put cost controls in place before usage spreads, not after you hit the wall.

The AI vendor default era is over. Here is what business owners should do instead.

For three years, most businesses defaulted to OpenAI. In May 2026, Anthropic overtook OpenAI in US business AI spending for the first time, according to the Ramp AI Index. DeepSeek went from near zero to the fastest-growing measured vendor in the same period. The AI vendor market has fragmented, switching costs are near zero, and the decision is no longer about picking the right tool. It is about building so you can change your mind.

Make better AI decisions, starting with one call.

Book a free AI Fit Call. We will tell you what to use, what to avoid, and where to start. No jargon, no pressure.