Microsoft’s custom MAI models take on GitHub Copilot and Excel

For years, the conventional wisdom has been simple: if you want an AI agent to handle complex work, you have to pay up for massive, expensive frontier models.

Microsoft's Superintelligence team is currently proving that rule wrong inside its own product stack.

Back at Build in June, Microsoft teased its "hill-climbing machine"—an integrated flywheel connecting real product data, models, and specialized evaluation harnesses. Today, the company revealed the first real-world fruits of that approach: specialized MAI models powering live agentic workloads inside GitHub Copilot and Microsoft Excel.

Smaller footprints, higher retention

The first model, dubbed MAI-Code-1-Flash, has actually been running quietly inside GitHub Copilot since June, serving millions of developers daily.

The production numbers tell an impressive story. In VS Code, MAI-Code-1-Flash posted an approximately 10% higher code acceptance rate than both GPT-5.4 Mini and Claude Haiku 4.5. Even better, developers were 6% more likely to return across multiple days compared to GPT-5.4 Mini, and 11% more likely than with Claude Haiku 4.5.

It managed all this while being significantly leaner. The model logged a 10% lower median token usage than its rivals, even as developers initiated more conversation turns.

Then came the Excel experiment. Microsoft wanted to test whether an agent trained on code could jump into general knowledge work. Instead of training from scratch, researchers took the MAI-Code-1-Flash checkpoint and dropped it into an Excel reinforcement learning environment to master spreadsheets and tool calling.

According to production traffic telemetry, the specialized Excel model now matches the output quality of GPT-5.6 on the most common spreadsheet tasks.

Crucially, it also changes the infrastructure math. The model runs smoothly on both Nvidia A100 and H100 class GPUs, meaning Microsoft doesn't need to monopolize scarce next-generation hardware just to run agentic spreadsheet formulas. Microsoft confirmed it is now taking this exact hill-climbing playbook to Copilot Chat, Outlook, and PowerPoint.

Why it matters

For everyday developers and office workers, this means snappier, more accurate assistants that actually understand the software they live inside without burning through unnecessary tokens or lagging under heavy inference loads.

For the wider industry, this is Microsoft signaling a major shift in how AI products will be served at scale. Instead of funneling all customer queries into giant, expensive models like GPT-5.6, owning the entire stack—from the UI and execution harness to real user feedback—allows companies to train hyper-efficient smaller models that punch far above their weight.

We think this approach is deeply pragmatic. Repurposing a coding base model for spreadsheet logic is a clever shortcut, and enabling these models to run on older A100 and H100 clusters relieves massive supply chain pressure on Microsoft's data centers. The lingering question is whether these specialized models can handle messy, out-of-distribution corner cases as gracefully as a raw frontier model. But for everyday enterprise workflows, hyper-specialized efficiency looks like the real path to profitable AI.