Microsoft MAI-Image-2 in Playground: What Microsoft's Own Image Model Can — and Can't — Do

Microsoft MAI-Image-2 is Microsoft AI's in-house text-to-image model — a model that turns a written prompt into an image — and you can still try it in Microsoft's MAI Playground, a limited preview for experimental MAI models. Microsoft says it debuted as the third-ranked model family on Arena's text-to-image leaderboard, a claim from Microsoft itself rather than an independent test, and that it generates images at least twice as fast as its predecessor. One correction matters first: MAI-Image-2 did not launch on 12 November 2026. It first appeared in the Playground on 19 March 2026, and a 12 November article about it is a recap, not a new release.

What MAI-Image-2 is

For years, Microsoft generated images in Bing Image Creator and Copilot almost entirely with OpenAI's models. MAI-Image-2 was the moment Microsoft's own team stepped forward with a competitive alternative. It was built by the Microsoft AI Superintelligence team, led by Microsoft AI chief executive Mustafa Suleyman, after that team was reorganised in March 2026 to focus on frontier models. Microsoft wanted its own image generator, trained and controlled in-house, so its products would no longer depend entirely on a partner for one of their most visible features.

Microsoft says it developed the model with input from photographers, designers and visual storytellers. That shaped its priorities: natural-looking photographs, accurate skin tones, realistic lighting, clear text inside images, and detailed or surreal scenes. It was aimed at creative professionals — people making thumbnails, posters and presentation images. In our flagship AI model comparison, the pattern is familiar: every major lab now wants its own image model. Its predecessor, MAI-Image-1, launched in October 2025 at around ninth on the same Arena leaderboard.

What the 12 November date really means

A new article about MAI-Image-2 really did appear on 12 November 2026, describing it as newly introduced in the Playground, with a rollout to Copilot and Bing Image Creator beginning and wider developer access through Microsoft Foundry still to come. But those details — the Arena ranking, the WPP early-adopter example, the Playground preview — are the same ones Microsoft announced in March. Reports from March 2026, including Winbuzzer on 20 March, record the same launch, and a formal Foundry release followed on 2 April 2026, when Microsoft announced MAI-Transcribe-1, MAI-Voice-1 and MAI-Image-2 together. The 12 November item is therefore a republication of an eight-month-old launch, not a new model or capability. MAI-Image-2 is still a real, usable model, and still worth explaining — but it is not November news. The genuinely new chapters are MAI-Image-2.5 (2 June 2026) and MAI-Image-2.7 (27 October 2026), covered below.

A photographer holding a camera in a bright studio, reviewing a portrait on the camera screen
AI-generated illustration (generated 15 Nov 2026)


What you can try in the Playground today

The most direct way to meet MAI-Image-2 is the MAI Playground, Microsoft's public testing environment. Microsoft describes it as a limited preview and warns that AI can make mistakes, so treat it as a place to experiment, not a production tool. You pick the image model, type a prompt, and get an image back. Beyond the Playground, Microsoft said at launch that the model was beginning to roll out to Copilot and Bing Image Creator, and later to PowerPoint. A rollout announcement is not the same as availability for every account — reports from March noted it had not yet appeared in Copilot on launch day — so check what your own apps actually offer. For developers, API access went to selected customers at launch, with WPP named as an early adopter, and broader Foundry access was described as coming soon. A good first test is a prompt with text in it, such as a poster, because in-image text was a stated selling point.

The limits that matter

MAI-Image-2 arrived with tight limits, and they matter more than its ranking. Hands-on reports from March 2026 describe a model strong on quality but deliberately restricted in use.

LimitWhat was reported
Aspect ratio1:1 square output only — no landscape, portrait or custom ratios
Daily cap15 images per day, then locked out for 24 hours
Cooldown30 seconds between generations
Content filteringAggressive filtering, reported as stricter than Google Imagen and OpenAI DALL-E
EditingText-to-image only — no image-to-image, inpainting, outpainting or reference-image support

Fifteen square images a day is enough to test ideas, but not a pipeline for a team producing dozens of images daily. The square-only output is a real constraint for social media, where portrait and landscape dominate. Reviewers in March reported that the model outperformed GPT-Image on image quality and text rendering in their tests, even though GPT-Image ranked higher on Arena. Treat it as a preview to evaluate, not a finished production service. Microsoft has not published a public Playground price, and these limits come from hands-on reports, not from Microsoft's announcement.

A designer editing a photograph on a large monitor in a modern studio
AI-generated illustration (generated 15 Nov 2026)


Pricing and the successors

For developers, the Foundry prices reported in April 2026 were $5 per 1 million tokens for text input and $33 per 1 million tokens for image output — API prices, not Playground prices. The business case was about independence as much as price: Microsoft reduced its reliance on OpenAI for Bing and Copilot imagery. What Microsoft did not announce was a general-availability date or a full rollout timetable.

On 2 June 2026, at Microsoft Build 2026, Microsoft announced seven new MAI models, including MAI-Image-2.5 and a Flash variant. Microsoft said MAI-Image-2.5 ranked second on Arena image editing, ahead of Nano Banana 2, and third for text-to-image, gaining 75 Elo points overall against MAI-Image-2 and 107 on text rendering. Unlike MAI-Image-2, it both creates images and performs precise localised edits, and it rolled out to PowerPoint and OneDrive. Then, on 27 October 2026, Microsoft announced MAI-Image-2.7 as an editing preview: second on Artificial Analysis, per Microsoft, and third on Arena editing, built to restyle, relight, fix text in, or restore an existing image. Microsoft said text-to-image for 2.7 was coming soon, with no API or pricing announced. The choice is simple: generate with 2.5, edit with 2.7. Canva's own AI image model shows the same industry pattern — every major platform now wants its own generator.

Printed photographs spread across a desk beside a laptop in warm light
AI-generated illustration (generated 15 Nov 2026)


How it compares with rivals

Microsoft's top-three claim needs careful reading. Arena rankings come from anonymous human-preference votes, not a standardised benchmark, so a model can rank third overall while beating a higher-ranked rival on a specific task. Rankings are also snapshots that shift as models and votes are added, so a March ranking is not a live result in November. Check the current Arena board before treating any rank as current. Kling 4.0 AI video generation shows how fast neighbouring AI video is moving, and Google Gemini 4 Carbon shows Google pushing its own flagships at the same time. MAI-Image-2's real achievement was a proof point: Microsoft could build an in-house image model reviewers took seriously, then iterate quickly to 2.5 and 2.7.

Key Takeaways

  • MAI-Image-2 debuted in the MAI Playground on 19 March 2026 — the 12 November 2026 article about it is a recap, not a new launch.
  • Microsoft says it ranked as the #3 model family on Arena's text-to-image leaderboard; that is Microsoft's own claim, based on human-preference votes.
  • It is text-to-image only, built for photorealism, accurate in-image text and detailed scenes.
  • Reported limits are strict: 1:1 square output only, 15 images a day, a 30-second cooldown, and aggressive content filtering.
  • Foundry API pricing reported in April 2026 was $5 per 1 million text-input tokens and $33 per 1 million image-output tokens; no public Playground price has been stated.
  • Newer versions have superseded it: MAI-Image-2.5 (2 June 2026) adds editing, and MAI-Image-2.7 (27 Oct 2026) is an editing-first preview.

Conclusion

Microsoft MAI-Image-2 deserves to be remembered as a turning point, not mistaken for November news. In March 2026 it showed Microsoft could build its own competitive image generator, earn a top-three Arena ranking on its own account, and ship it to the Playground, Copilot and Bing Image Creator. Its limits kept it a preview in practice, and its successors have already taken the family further. If you want to try Microsoft's image models today, start in the Playground, test the in-image text Microsoft highlights, and compare the result with MAI-Image-2.5 before building anything on the older model.

Sources

Comments