Microsoft released two in-house AI models into public preview Wednesday, marking its most aggressive push yet to reduce dependence on OpenAI. The models, MAI-Image-2.5-Pro and MAI-Voice-2-Flash, deliver substantial cost savings while maintaining competitive performance.

MAI-Image-2.5-Pro represents Microsoft's highest-fidelity image generation capability to date. The model handles complex visual tasks with improved quality compared to earlier iterations. MAI-Voice-2-Flash targets enterprise speech applications at scale, optimized for high-volume workloads where inference costs dominate budgets.

The cost advantage proves decisive. Microsoft's production data shows the new models cut expenses by up to 89 percent compared to OpenAI's offerings. This gap widens particularly for voice applications, where enterprise customers process billions of tokens monthly. For image generation, the savings remain substantial while quality metrics match or exceed competitor benchmarks.

The timing reflects broader strategic shifts within Microsoft. The company has invested heavily in OpenAI but faces pressure to reduce per-token costs as AI adoption scales across Office, Copilot, and Azure services. Building proprietary models allows Microsoft to capture the margin difference and negotiate from stronger positions with OpenAI.

These releases follow a pattern. Microsoft published Phi models targeting smaller, efficient deployments. It integrated Copilot deeply into Windows and Office applications. It built Copilot Stack specifically for enterprise customers. Each move reduces reliance on frontier models from external partners.

The public preview invites developers and enterprises to test the models in real workloads. Production data backing the cost claims carries weight. Microsoft measured actual inference performance, not theoretical benchmarks. This contrasts with typical AI launches where claimed improvements often fail to materialize in practice.

For enterprises, the choice shifts from pure capability to capability-per-dollar. If MAI models deliver