Microsoft AI released MAI-Image-2.6 on September 4, 2026, calling it the lab’s strongest image model. A companion model, MAI-Image-2.6-Flash, is aimed at latency-sensitive production queues. Both are in public preview on Microsoft Foundry and available to try in the MAI Playground.

The company said MAI-Image-2.6 ranks No. 2 for text-to-image and image editing on Arena as of September 4. On Artificial Analysis, it ranks No. 2 for text-to-image and No. 1 for image editing. Microsoft’s pitch is price per quality, not a clean sweep of the boards: the post says the model has the “best price-per-Elo performance in the world.”

Flash is the volume SKU. Microsoft said it generates images 2.8 times faster than OpenAI’s GPT-Image-2-Medium and is 72 percent more efficient. Developers pick 2.6 when they want maximum precision and Flash when they need throughput.

Editing, grounding, and 1.5K frames

Both models add three production features the earlier MAI-Image-2 line did not lead with. Multi-reference editing can pull people, products, styles, and scenes from different stills into one frame. Web grounding can fetch current facts for a prompt instead of inventing a scene from memory. Dynamic aspect ratios pick a layout automatically, with output up to 1.5K resolution.

The release lands a day after the same lab shipped MAI-Transcribe-2 at 10 cents an hour. Image-2.6 extends that pattern: a first-party MAI model priced and packaged to sit next to OpenAI and Google endpoints inside Foundry, rather than only inside Copilot.

Decoded Take

Microsoft is no longer treating first-party image models as a research sidebar. Image-2.6 is a Foundry SKU with a Fast twin, a leaderboard footnote, and a cost claim that matters more than the No. 2 Arena rank. The feature list (multi-reference, web grounding, variable aspect) is what design teams actually ask for when they stop demoing and start generating catalog shots. The risk is the usual Microsoft sandwich. A sharp MAI endpoint can still get folded into a Copilot bundle with less transparent metering. Watch whether independent Arena and Artificial Analysis boards hold the September 4 ranks after a week of voting, whether Flash stays cheaper than GPT-Image-2-Medium on real invoices, and whether web grounding cites sources or just paints over them.