Microsoft AI on September 4, 2026 opened public preview access in Microsoft Foundry to MAI-Image-2.6, which it calls its strongest image model yet, and introduced MAI-Image-2.6-Flash for latency-sensitive, high-throughput workloads. Both models support text-to-image generation and image editing with multi-image reference inputs, web grounding, and dynamic aspect ratios up to about 1.5K resolution.
In its official announcement, Microsoft said MAI-Image-2.6 ranked No. 2 for both text-to-image and image editing on Arena as of September 4, and No. 2 for text-to-image plus No. 1 for image editing on Artificial Analysis. The company positioned Flash as delivering comparable quality with claimed generation about 2.8x faster than GPT-Image-2-Medium and 72% greater efficiency—figures that remain Microsoft-reported rather than independently benchmarked in the launch post.
Independent coverage from Runtime Wire corroborated the Foundry public-preview timing and noted Arena score estimates placing MAI-Image-2.6 near the top of crowd-ranked text-to-image boards, still trailing OpenAI GPT Image 2 while competing closely with other frontier image systems. Runtime Wire also summarized Microsoft token pricing cited for the models, including higher rates for the full MAI-Image-2.6 stack and lower rates for Flash; per-image cost depends on resolution and token use.
Microsoft’s accompanying model materials describe a diffusion-based architecture with roughly 20 billion non-embedding parameters, image outputs capped near 1536×1536 total pixels, and availability through MAI Playground plus Azure AI Foundry. The September 4 event expands Foundry from private preview and adds the Flash variant after MAI-Image-2.6 first appeared in Arena testing in August.
For production teams, the practical takeaway is choice: maximum precision via MAI-Image-2.6, or faster, cheaper generation via Flash, both under Microsoft’s own model stack inside Foundry rather than only third-party image APIs.