News

Microsoft Expands In-House AI Push with New MAI Models for Developers

• By Chris Paoli
• April 03, 2026

Microsoft announced three new artificial intelligence models on Thursday, marking an expansion of its MAI (Microsoft AI) portfolio with offerings designed to compete directly on price and performance with others, including Google.

Deepening investment in proprietary AI with the release under its MAI umbrella, targeting enterprise developers looking for alternatives to existing offerings from hyperscalers and model providers. The move signals a continued shift toward building competitive, first-party AI capabilities within Azure.

The newly announced models, MAI-Transcribe-1, MAI-Voice-1 and MAI-Image-2, are available via Microsoft Foundry and the MAI Playground, giving organizations tools for speech-to-text, voice generation and image creation. Positioned around both cost efficiency and performance, the models are designed to go head-to-head with similar technologies from rivals such as Google and other AI vendors.

Speech and Voice Capabilities Lead the Pack
MAI-Transcribe-1 aims to deliver speech-to-text transcription across 25 languages. According to the company, its batch transcription speed is 2.5 times faster than the existing Microsoft Azure Fast offering, while maintaining what the company describes as world-class quality in real-world environments.

The model starts at $0.36 per hour, positioning it as "the best price-performance of any large cloud provider," according to Microsoft.

MAI-Voice-1 can generate 60 seconds of audio in just one second. The model now includes the ability to create custom voices with only a few seconds of audio input, a feature Microsoft says will "transform how easily developers can build voice experiences and voice agents." Pricing for MAI-Voice-1 starts at $22 per 1 million characters.

Image Generation Upgrade
Perhaps the most notable of the three releases is MAI-Image-2, which Microsoft says has already achieved top-three status on the Arena.ai leaderboard. The model delivers generation times that are at least twice as fast as previous versions.

The image generation model was developed with input from photographers, designers and visual storytellers, with a focus on natural lighting, accurate skin tones and texture, and clear in-image text for diagrams, layouts and graphics. Microsoft is rolling out MAI-Image-2 in phases across Bing and PowerPoint, in addition to Copilot. Pricing starts at $5 per 1 million tokens for text input and $33 per 1 million tokens for image output.

Responsible AI at the Core
Microsoft said that all three models were developed with its commitment to safe and responsible AI, undergoing development, testing and rigorous red-teaming before release. Through Microsoft Foundry, developers receive built-in guardrails, governance and enterprise-grade controls designed to support safe, compliant deployment at scale.

The models are available immediately for developers through Microsoft Foundry, with U.S.-based users able to test them in the MAI Playground. Developers without Foundry access can request access through a form on the Microsoft AI Web site.

Agent Evaluation Tools Reach General Availability
In a related announcement, Microsoft said Agent Evaluation in Microsoft Copilot Studio has reached general availability, providing enterprise customers with new capabilities to assess AI agent behavior at scale.

The evaluation feature addresses what Microsoft says is a critical challenge as organizations move AI agents into production environments. According to the company, manual testing can't scale to handle the hundreds or thousands of interactions that production agents manage, making automated evaluation essential for ensuring consistent, safe behavior.

Agent Evaluation is built directly into Copilot Studio and provides an end-to-end workflow for creating test cases, running evaluations and reviewing results without requiring code. The tool produces versioned and auditable results that include the test set used, the user profile that ran it, the date and duration, and results from each grader for every test case.

"As agents move into production, evaluations help take each build from experimentation to a reliable system," Microsoft stated in its announcement. The company positioned the tool as helping answer a critical question for production deployments: "Can we trust this agent to behave correctly, consistently, and safely -- every time?"

The evaluation system supports identity-based testing, allowing teams to evaluate agents under specific user profiles with the same knowledge sources, tools and connectors they would access in production. For organizations running continuous integration and delivery pipelines, Agent Evaluation is also available via API, enabling teams to integrate assessments directly into existing deployment workflows