Microsoft Copilot: Why the AI Inside It Is Getting Quietly Cheaper
Agentic AIMicrosoft’s AI head has stated the company aims to “reduce and ultimately eliminate” payments to OpenAI and Anthropic, describing them as excessive. The shift is already live — and enterprise Copilot subscribers are the ones absorbing the consequence.
The Cost Architecture Behind Copilot
Microsoft has been paying OpenAI and Anthropic billions annually to power the AI layer across its productivity suite. That arrangement made sense when Microsoft lacked competitive in-house models. With the development of its proprietary MAI model line — anchored by MAI-Thinking 1, its flagship reasoning model — the calculus has shifted.
The transition is already underway. MAI models now handle “tens of thousands of requests per week” in Excel and Outlook alone, with GitHub Copilot and Teams listed as subsequent targets. Microsoft did not announce the change to customers. Users running Copilot on flat-rate enterprise licenses are receiving a different AI than the one that informed their procurement decision.
What the Benchmark Reality Looks Like
The commercial rationale is clear. The performance trade-off is less so. MAI-Thinking 1, the reasoning model being deployed as a replacement, underperformed against published benchmarks, trailing both OpenAI and Anthropic competitors and performing “roughly on par with DeepSeek V3.2.” DeepSeek V3.2 is a capable open-weight model — but it is not what enterprise customers expect from a Microsoft 365 Copilot subscription marketed on GPT-4-class capability.
The pricing structure being considered makes the dynamic explicit: MAI becomes the default at current rates, while access to OpenAI or Anthropic models becomes a paid add-on. Microsoft would initially absorb vendor costs, then eventually pass them through. Flat-rate subscribers would receive the lower-capability tier without a corresponding price reduction.
An additional complication: Microsoft marketed its MAI training advantage as using “commercially licensed data,” but reporting indicates the actual corpus relied heavily on Common Crawl web data — the same unlicensed source at the center of ongoing legal disputes in the AI industry.
The So What
For organizations running Copilot at scale, this is a procurement risk that standard enterprise software contracts don’t account for. AI capability agreements negotiated around GPT-4o or Claude-class performance may now deliver MAI-class output without recourse or renegotiation. The signal from Microsoft’s leadership — framing third-party model payments as something to eliminate rather than optimize — suggests this is a strategic trajectory, not a temporary cost experiment. Teams that depend on reproducible AI quality in production workflows, from legal document drafting to code generation in GitHub Copilot, should be auditing which capabilities are model-dependent before the next renewal cycle. Related: the emerging economics of enterprise AI procurement.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →