Microsoft Adopts MAI Models to Reduce Token Costs

Bloomberg, cited by Gizmodo, reported that Microsoft is using its own MAI models to complete "tens of thousands" of AI prompts each week in Excel and Outlook instead of relying exclusively on third-party models, per an unnamed source, Gizmodo reports. Gizmodo adds that this represents a small portion of Microsoft's overall AI usage and that high-volume services such as Copilot still drive the majority of token consumption. Microsoft declined to comment on the Bloomberg report, Gizmodo reports.

Gizmodo reports Microsoft recently announced the model MAI-Thinking-1, which Microsoft described as built for "high efficiency and performance, but importantly, at a low-token cost," and lists it as a 35 billion active-parameter model with a 256K context window. Gizmodo also reports Microsoft rolled out additional image, transcription, voice, and coding models alongside MAI-Thinking-1.

Industry-pattern observations: As model inference costs scale with both usage and context length, larger enterprises are increasingly weighing internal model deployment or cheaper mid-sized models to reduce per-query expense. Observers following the sector have noted a growing trade-off space where mid-sized, optimized models can match task-level performance for many product tasks while substantially lowering token bills.

Signals observers should follow include vendor token pricing changes, published cost-per-token benchmarks for comparable tasks, third-party model adoption rates inside major SaaS products, and any Microsoft statements or blog posts clarifying production rollouts or cost-per-query metrics. Bloomberg sourced the initial usage claim to an unnamed source; Microsoft has not provided a public explanation of the reported routing decisions, Gizmodo reports.

The practical takeaway for ML engineers and platform teams is straightforward, token economy is becoming a first-order constraint for large-scale, product-facing LLM features. Teams designing in-app assistants or background generation should treat token price and model efficiency as architectures-level trade-offs, not only accuracy metrics.

For engineering teams, the practical actions are to benchmark both accuracy and cost for target prompts, include token-cost forecasts in product planning, and design runtime routing that can fall back to cheaper in-house or mid-sized models when acceptable. This story reinforces that token economics, not only raw model capability, determine which models run at scale in production.