Microsoft Pivot: Shifting Copilot to Cheaper MAI Models and Launching Microsoft Foundry
Microsoft has initiated a sweeping operational pivot in how it powers its massive Copilot ecosystem, introducing a multi-model orchestration strategy aimed at dramatically lowering the massive compute costs of enterprise AI.12 According to Microsoft Chairman and CEO Satya Nadella, the tech giant is systematically migrating high-volume, everyday tasks across GitHub Copilot, Excel, and Outlook away from expensive frontier models and onto its proprietary MAI (Microsoft AI) family.
Rather than relying on general-purpose frontier giants like OpenAI's GPT or Anthropic's Claude for every request, Microsoft is deploying specialized, home-grown models optimized for specific product workflows, reserving frontier intelligence only for the most demanding, complex queries.
Specialization Over Scale: MAI Outperforms Frontier Models in testing
Microsoft's early production data shows that highly targeted, smaller models can outperform massive general-purpose flagships on specific tasks while operating at a fraction of the token cost:
- Excel workflows: Microsoft’s specialized MAI model performs on par with OpenAI's flagship GPT-5.6 for common spreadsheet tasks.
- GitHub Copilot: The newly deployed MAI-Code-1-Flash model recorded an approximately 10% higher code acceptance rate than GPT-5.4 Mini and Claude Haiku 4.5 in internal testing. Crucially, it achieved this while utilizing 10% fewer median tokens, representing a double-digit improvement in both developer productivity and inference efficiency.
Instead of training a single, trillion-parameter model to perform all tasks, Microsoft’s strategy focuses on training specialized models inside simulated environments tailored to the target product. This allows an Excel model to master spreadsheet logic, and a GitHub Copilot model to learn developer workflows, directly in the software environments where they reside.
"Hill Climbing" and Microsoft Foundry
Microsoft calls its continuous optimization process "hill climbing." Rather than evaluating its models on static academic benchmarks, the company measures success based on whether the models successfully complete the specific tasks and workflows that customers actually care about.
To help enterprise customers replicate this cost-saving architecture, Microsoft is launching Microsoft Foundry. This framework allows businesses to build, evaluate, and optimize their own autonomous AI agents using their private data, custom workflows, and tailored evaluation systems.
While this pivot represents a major step toward reducing Microsoft's financial exposure to high-cost API partners, Nadella emphasized that OpenAI and Anthropic are not disappearing. Instead, they will remain integrated as the high-tier "reasoning engine" within Microsoft's broader orchestration system, handling complex edge-case tasks while MAI models handle the high-volume, everyday workloads.
Verbatim Quotes
"Instead of relying on the most powerful and expensive frontier model for every request, Microsoft wants to use the right model for each job. The goal is simple: maintain or improve AI performance while using fewer tokens and reducing the enormous computing cost of serving millions of users." — WION, Microsoft has a new plan for AI: Use expensive frontier models only when needed
"'We can now take saturated frontier capabilities and deliver them at scale and at lower cost through models optimised for high-usage products, while continuing to use frontier models for frontier needs,' Nadella wrote... 'We are now seeing MAI models outperform general-purpose frontier models in many use cases while using a fraction of the tokens.'" — Satya Nadella, quoted in Satya Nadella says Microsoft’s MAI AI models outperform frontier AI in many use cases, cut costs
-
An instance of High-volume software platforms must migrate from general-purpose frontier models to specialized, proprietary micro-architectures. — Microsoft transitions its high-volume Copilot applications away from general-purpose APIs and onto specialized, in-house MAI models to lower running costs. ↩︎
-
An instance of Sustainable software margins require decomposing monolithic tasks into multi-model hierarchies. — Microsoft is migrating everyday Copilot requests away from general-purpose frontier giants to its own cheap, specialized micro-models to control costs. ↩︎