Cheaper and Better: What Microsoft's MAI Numbers Teach You About Model Routing
Microsoft published what it saved by moving workloads off third-party frontier models. The headline was 84–89% GPU cost cuts. The number that should change your thinking is a 26% rise in save rates.
Microsoft published what it saved by moving workloads off third-party frontier models. The headline was 84–89% GPU cost cuts. The number that should change your thinking is a 26% rise in save rates. What happened On 23 July 2026 , Microsoft AI put two in-house models into public preview: - MAI-Image-2.5-Pro — its highest-fidelity image model, aimed at detailed generation and editing with precise in-image text rendering . Pricing: $5 per million text input tokens, $8 per million image input tokens, $106 per million image output tokens . - MAI-Voice-2-Flash — a text-to-speech model for low-latency, high-volume voice work such as customer-service agents. 2× faster and 32% cheaper than MAI-Voice-2 , at $15 per million characters . Both are available through Azure / Microsoft Foundry, and free to try in the MAI Playground. Then came the part worth reading twice. Microsoft disclosed what happened when it replaced third-party models with its own inside real products: Product Model Published result --- --- --- Bing Image Creator MAI-Image-2.5 Now the default model — full substitution PowerPoint MAI-Image-2.5 GPU cost reduced up to 84% OneDrive MAI-Image-2.5 Save rates +26% , latency -25% Dynamics 365 Contact Center MAI-Voice-2-Flash GPU cost reduced up to 89% It also mentioned MAI-Transcribe-1.5 powering Dragon Copilot across 58 languages with a 50% relative reduction in transcription errors . (A small housekeeping note: several outlets dated this launch 28 July. Microsoft's own post says 23 July, and we're going with the primary.) Why it matters Vendors publish benchmark scores constantly and production economics almost never. That alone makes this disclosure useful — four specific, product-level results with numbers attached. But almost every write-up led with the 84% and 89% cost figures and skipped the row that actually matters. OneDrive: save rates up 26%, latency down about 25%. Save rate is a quality proxy. It measures users keeping the generated output rather than discarding it — which is a far better signal than any benchmark, because it's revealed preference on real work. So Microsoft is reporting that moving off a third-party frontier model made the feature cheaper, faster, and better received by users , simultaneously. That breaks the assumption almost every product team operates under: that moving off the frontier model is a quality sacrifice you accept in exchange for margin. Sometimes it isn't a sacrifice at all. For a CEO, that's the number to bring to a margin conversation — a public, citable example of 85% serving-cost reduction on specific workloads without a quality regression. For a CPO, it changes which features you're willing to put on the table. And for anyone whose AI feature margins are currently underwater, it reframes the problem from "we need cheaper inference" to "we're using a general-purpose model for a narrow job." Under the hood Why substitution can improve quality. A frontier model is optimised for breadth — it has to b