- Companies are now aggressively exploring multi-model routing strategies, directing simple queries to cheaper models and reserving premium ones for complex, high-value tasks.
Microsoft has begun a quiet but consequential overhaul of its AI strategy: replacing models from OpenAI and Anthropic with its own proprietary MAI models across key parts of its Copilot portfolio.
The move, already underway in Excel and Outlook, marks the beginning of what industry watchers are calling a defensive pivot — one driven less by technical superiority than by the brutal arithmetic of runaway AI costs.
Tens of thousands of Copilot prompts that would have previously been routed to OpenAI’s GPT models or Anthropic’s Claude are now being handled by Microsoft’s own in-house systems. While the volume still represents a fraction of total user queries, it signals a definitive and long-term strategy to wean the company off expensive third-party AI vendors.
The financial motivation behind the shift is unambiguous. Flat-rate subscriptions — the model that made Copilot accessible to millions — have turned out to be economically disastrous when heavy users enter the equation.
At Microsoft’s Build conference in June 2026, Mustafa Suleyman, the company’s head of AI, didn’t mince words about the problem.
“We pay a lot of money to Anthropic — so our goal is to reduce and ultimately eliminate that cost,” he told attendees. His bluntness reflected a growing frustration across the industry with what has come to be known as “tokenmaxxing” — the practice of consuming as many AI tokens as possible, often without regard for cost, under flat-rate enterprise plans.
The scale of the problem is staggering. According to recent research cited by multiple outlets, a single heavy user on a premium $200 monthly AI plan can generate up to $14,000 in actual token costs, forcing AI labs to heavily subsidize usage — a model that is manifestly unsustainable at scale .
Uber’s wake-up call
Microsoft is far from alone in confronting this reality. Uber became the cautionary tale of 2026 when reports surfaced that the company had burned through its entire annual AI budget in just four months, driven largely by surging developer adoption of automated coding tools like Claude Code. Monthly costs per engineer averaged between $150 and $250, with power users running up bills of $500 to $2,000 each.
Uber has since imposed a $1,500 monthly cap per employee per AI tool — a blunt but increasingly common response. As CNBC recently framed it, corporations are now confronting an uncomfortable trade-off: “tokens or humans,” with CFOs weighing AI spend against future headcount.
The broader picture is sobering. Gartner projects AI agent software spending will reach $207 billion in 2026, up 139 per cent from the prior year — yet many enterprises are already discovering that AI is costing more than the people it was meant to replace.
MAI models: Microsoft’s homemade answer
At its Build conference, Microsoft debuted seven in-house MAI models spanning reasoning, coding, image generation, speech, and transcription. The headline act was MAI-Thinking 1, the company’s first reasoning model, which Microsoft claimed could compete with Anthropic’s flagship Sonnet 4.6 and Opus 4.6 in coding benchmarks.
Independent assessments, however, have painted a more complicated picture. MAI-Thinking 1 scored a strong 97 per cent on AIME 2025, a competition-level mathematics benchmark, and demonstrated competitive performance on GPQA and SWE-Bench Verified. But broader evaluations suggest the model trails top-tier US rivals by a meaningful margin, with some analysts characterising its overall performance as closer to open-source alternatives like DeepSeek V3.2.
The disparity raises a pointed question: is Microsoft trading capability for cost savings? For routine tasks in Excel and Outlook — summarising spreadsheets, drafting emails, generating basic analysis — the answer may be that MAI models are “good enough.”
The company is betting that most enterprise users won’t notice the difference, especially when the alternative is a price hike or usage cap.
The Chinese price shock
Microsoft’s pivot also unfolds against a geopolitical dimension reshaping the global AI market. While premium Western models from OpenAI and Anthropic charge upward of $4 per million tokens, Chinese providers have slashed prices to as little as 18 cents per million tokens — a more than 20x differential — while rapidly closing the capability gap .
According to CitiBank Research cited by Reuters, Chinese models delivered 4.12 trillion tokens in a single week in February 2026, compared to 2.94 trillion from American models — a reflection of both price sensitivity and the growing competitiveness of open-source and state-backed Chinese AI labs .
This price war has fundamentally altered the procurement calculus for enterprise buyers. As one Reuters headline put it: “Cheaper AI is better.”
Companies that once defaulted to OpenAI or Anthropic are now aggressively exploring multi-model routing strategies, directing simple queries to cheaper models and reserving premium ones for complex, high-value tasks.
What this means for Copilot users
For the millions of workers who rely on Copilot inside Excel, Outlook, and other Microsoft 365 applications, the shift to MAI models is likely to be invisible — at least at first. The company has been careful to route only selected, lower-complexity prompts to its in-house models, preserving premium third-party models for tasks that genuinely require them.
But the direction of travel is clear. Suleyman’s language — “reduce and ultimately eliminate” — leaves little room for ambiguity. Over time, Microsoft intends for its own models to handle an ever-larger share of Copilot workloads. The question is whether its models can keep pace with the furious rate of improvement coming from labs like Anthropic, OpenAI, and the increasingly formidable Chinese ecosystem.
