Updated on August 16, 2026
MAI-Image-2.6 is Microsoft AI's in-house image generation model, the latest release in the MAI-Image family. At launch it took second place on Arena's text-to-image leaderboard out of 77 systems, trailing only OpenAI's GPT-Image-2 and finishing ahead of entries from Google, Meta and xAI. Testing it costs nothing on Arena, with MAI Playground next in line and paid API access coming through Microsoft Foundry. Designers, marketing teams and online shops tired of garbled lettering in generated images are the obvious audience.
Text rendering improves by +91 Elo over MAI-Image-2.5, the biggest single jump in this release. You type a three-word tagline for a poster and it comes back spelled correctly, with clean spacing (anyone who has binned a mockup over melted letters will feel that one). Microsoft also reports more natural portraits and more polished product and photorealistic output for advertising and e-commerce.
There is more on the way. Working from several reference images, richer grounding in real data and finer control over reasoning, format and resolution have all been announced, though full documentation is still pending.
| Arena category | MAI-Image-2.5 rank | MAI-Image-2.6 rank |
|---|---|---|
| 3D imaging and modeling | 6th | 1st |
| Text rendering | 8th | 2nd |
| Product, branding and commercial design | 7th | 2nd |
| Cartoon, anime and fantasy | 8th | 2nd |
| Art | 4th | 2nd |
The model entered Arena's text-to-image board with 1336 Elo points, 45 behind GPT-Image-2 and 20 ahead of xAI's Grok Imagine Image 2.0. The family has come a long way, since MAI-Image-1 debuted in 9th place and versions 2 and 2.5 both stalled at third. In ten months, Microsoft turned a budget option into a title contender.
One caveat is worth keeping in mind. Arena relies on blind human votes scored through an Elo rating system, and 2.6 had gathered roughly twenty times fewer ballots than Nano Banana at entry, so positions can shift. Beating Meta's Muse Image and Google's models is still a serious calling card among the leading text-to-image models.
Trying the model is free. Arena lets you generate images and compare outputs blind at no cost, long enough to form your own opinion. No official pricing had been published for 2.6 at launch, and billing will run through Microsoft Foundry as it did for earlier family members.
The rollout follows three stages laid out by Microsoft.
Yes, for testing. The model can be tried at no cost on the Arena platform, where anyone can generate images and vote blind between two outputs. Professional API access will run through Microsoft Foundry and will be billed, but Microsoft had published neither a price list nor API identifiers at launch time.
Not officially yet. Microsoft has announced a rollout to Foundry and then to its other products and services, without a detailed timeline. Earlier MAI-Image models already power surfaces such as Bing Image Creator, so 2.6 will likely follow, but no date has been confirmed for Copilot so far.
No, not on the overall board. At its Arena debut, MAI-Image-2.6 scored 1336 points against 1381 for GPT-Image-2, a 45-point gap. Microsoft's model does take first place in the 3D imaging and modeling category, and climbs to second for text rendering and commercial design.
Closed, entirely. MAI-Image models ship as closed weights and can only be reached through Microsoft surfaces, meaning Arena for public testing, MAI Playground, the Microsoft Foundry API or in-house products like Bing Image Creator. No downloadable version exists, unlike open families such as Stable Diffusion.
Verdict: Posters, packaging, product pages, anywhere a misspelled word ruins the shot, MAI-Image-2.6 deserves a test run on Arena, and teams already building on Microsoft Foundry get a first-party image engine that finally competes at the front of the pack.
