Cost and model control shape Microsoft’s new MAI lineup

At Ignite, Microsoft expands MAI across coding, images, voice, and security, with customization options and staggered availability.

Keeping agents running continuously, with models suited to their work and budget, is the connecting thread in the MAI announcements at Ignite. Microsoft is expanding its lineup across software development, image creation, voice, and cybersecurity. The company is also introducing licensed weights, allowing businesses to customize models within their own Azure environments. The announcement includes Maia 200, Microsoft’s second-generation AI accelerator, now running its largest models in production. Availability varies by product: some are already integrated into applications, while others remain in preview or are scheduled to arrive in the coming weeks. Targeted Image Editing and a Canvas Workspace

The MAI Image 2.7 and 2.7 Flash models cover image generation and editing. Highlighted capabilities include localized changes to color, material, or position, consistency across successive edits, and transparent backgrounds. In MAI Labs, the new Canvas workspace uses the Flash version to start from text or a sketch, then select an area to replace, move, or transform.

The 2.7 version leads the Artificial Analysis image-editing leaderboard, with an Elo score of 1,215 across 5,478 votes. Its lead over GPT Image 2.6 Sunburst at the maximum setting is four points, with overlapping uncertainty intervals. This result reflects preferences on the editing tasks evaluated rather than establishing superiority across all visual applications. Both models are available in public preview in Microsoft Foundry, with indicative costs of approximately $35 per 1,000 images for version 2.7 and $10 for Flash. Specializing in Everyday Coding and Structured Decisions

The two coding models divide work according to its scope. Code 2 targets deep debugging, changes across multiple files, and interfaces; Code 2 Flash focuses on explanations, targeted fixes, and frequent iterations. Microsoft specifies that these models are not yet optimized for fundamental scientific research or very long-horizon agentic work.

Averaged across four benchmarks, the company reports that Code 2 delivers 98% of GPT-6.1 Sol’s performance at 40% of its cost, and 96% of Claude Opus 5.5’s performance at 5% of the cost. These comparisons use maximum reasoning effort and MAI’s 50% promotional discount. Three tests ran on the internal MSBench platform using GitHub Copilot CLI without subagents; FrontierCode data comes from Artificial Analysis. Both models are announced as available in GitHub Copilot, with Flash also in limited-access preview in Foundry.

A derivative of Code 2 Flash, MAI Work Flash, also serves several Microsoft 365 and FabricIQ products, from office tasks to DAX query generation for Power BI.

For decisions embedded in applications, Decision 1.1 serves a different purpose: model routing, classification, prioritization, or verification, choosing among actions defined by the developer. It now accepts images and provides confidence scores to set thresholds for taking action or escalating. Available in Foundry, OpenRouter, and Vercel, it starts at $0.042 per million input tokens in Foundry, with no charge for outputs. Voice and Cybersecurity Have Their Own Deployment Conditions

The MAI Realtime 1 model handles voice conversations, including interruptions, tone adjustments, and connections to enterprise tools. Base pricing starts at $0.01 per audio minute, with reasoning billed separately at $1 per million input tokens and $5 per million output tokens. Testing is available through MAI Labs or a limited-access preview in Foundry. Integrations are also announced for Dynamics 365 Contact Center and Dragon Copilot, where it will support simulated patient conversations.

Transcription follows a separate path: Dragon Copilot Transcribe, built on MAI Transcribe 2, is already deployed in Dragon Copilot. Microsoft attributes 10 million transcribed patient encounters per month to the service and reports that error rates have been halved in multilingual workflows.

In security, MAI Cyber 2 Flash achieves a 92.3% success rate versus 91.5% for GPT-6 Astra on CyberGym-E2E, according to Microsoft. The evaluation excludes 92 tasks overlapping with training data for every model, leaving 813 tasks. The calculated offline evaluation cost is $0.45 per task versus $34.38, using the selected list prices. The model will be exclusive to Microsoft security products; replacement of the existing MAI profile in MDASH is scheduled to begin in mid-December for eligible customers. Customizable Weights Within the Company’s Azure Environment

The MAI Licensed Weights offering is intended to give businesses control over deployment, access, and upgrades for customized models within their own Azure environments. Microsoft says customer data and adaptations remain theirs and are used to train Microsoft’s models only with explicit consent. Commercial access is coming soon through Frontier Company engagements.

Published examples concern adaptations for specific tasks. At Alaska Airlines, a version of Code 2 Flash tuned to classify maintenance reports improves accuracy from 81.2% to 91.1%; qualified employees retain responsibility for final decisions. At Woven by Toyota, a model adapted for coding-standard fixes produces ready-to-merge changes in 44% of cases where the models used in production required human intervention.