MAI-Image-2.6 starts near the top
Microsoft AI announced MAI-Image-2.6 on August 11, 2026, presenting it as the next step in its image-generation model family. The current Text-to-Image Arena leaderboard places the preview model second with a score of 1,336 (+/-11), based on 3,488 votes.
OpenAI’s GPT Image 2 (Medium) remains first at 1,381 +/-5. That leaves Microsoft 45 points behind the leader. MAI-Image-2.6 is also 20 points ahead of Grok Imagine Image 2.0 (Low), which sits third at 1,316 +/-12 and is marked preliminary in the Arena data.
The ranking is not a laboratory benchmark and should not be read as a universal measure of image quality. Arena scores reflect anonymous user preferences in head-to-head comparisons. They are useful precisely because they capture what people choose, but they do not tell us everything about latency, reliability, licensing, safety filters or production cost.
A major jump over MAI-Image-2.5
The more revealing comparison is inside Microsoft’s own family. MAI-Image-2.5 is currently tenth overall at 1,256 points. MAI-Image-2.6 therefore improves by 79 Elo points, matching Microsoft’s own announcement. Microsoft also reports a 91-point improvement in text rendering and gains in every measured Arena category.
Those gains matter because image generators are no longer judged only by whether they can produce an attractive landscape. Text inside images, portraits, product photography, 3D scenes and commercial compositions are now everyday workloads. A model that can place readable labels or maintain a coherent product design is more useful to a European retailer or small agency than one that merely produces a visually striking first attempt.
Microsoft highlights stronger text rendering, portraits and 3D imagery, together with more polished product, branding, cinematic and photorealistic outputs. It also says that the new model is designed to work with multiple references and offers more control over reasoning, format and resolution. These are promising claims, but they are not yet a substitute for independent tests of editing consistency, prompt adherence and failure rates.
The price comparison is unusually important
At launch, there is no public per-image API price for MAI-Image-2.6. Microsoft says the model is available for text-to-image testing on Arena, is coming to MAI Playground later this week, and will roll out soon across Microsoft Foundry and other products. That gives users a route to try it, but not yet a predictable production budget.
For comparison, the Arena listing currently shows a reference price of $0.046 per image for GPT Image 2 (Medium). Using the European Central Bank’s August 11 reference rate of $1 = €0.8668, that is approximately €0.040 per image. The Arena listing does not provide a public price for MAI-Image-2.6, so it would be misleading to call Microsoft’s model cheaper or more expensive at this stage.
Google’s Gemini 2.5 Flash Image, better known as Nano Banana in some Google interfaces, has a published standard-tier output price of $0.039 per image, or about €0.034 at the same exchange rate. Google’s batch price is $0.0195 per image, approximately €0.017, while the priority tier is $0.0702, approximately €0.061. Google bills the image output using a token-equivalent calculation: an image up to 1,024 by 1,024 pixels consumes 1,290 output tokens at the published rate.
That produces a useful baseline:
- MAI-Image-2.6: no public API price yet; Playground and Foundry access are announced as coming soon.
- GPT Image 2 (Medium): $0.046 / approximately €0.040 per image in the current Arena price data.
- Gemini 2.5 Flash Image: $0.039 / approximately €0.034 per image on the standard tier; $0.0195 / approximately €0.017 in batch.
The comparison is not perfectly like-for-like. The Arena price field and Google’s API tariff may refer to different service configurations, resolutions, limits or commercial conditions. VAT, regional Azure billing, platform fees and enterprise agreements can also change the final amount paid by an EU customer. Still, the figures show the hurdle Microsoft faces: a model that ranks second will attract serious attention, but developers will need a price at least in the same neighbourhood before they redesign a production pipeline around it.
What does this mean for European users?
MAI-Image-2.6 is not yet a generally available Microsoft Foundry endpoint, so its EU availability and regional deployment details remain to be confirmed. Microsoft’s existing cloud distribution gives it a credible path into European enterprise workflows, but “coming to Foundry” is not the same as confirming a specific EU region, data-residency option or contractual privacy terms.
That distinction matters under the GDPR. A business generating product mock-ups or campaign visuals should check where prompts and uploaded reference images are processed, how long they are retained, whether they are used for service improvement, and which Microsoft entity acts as processor. The EU AI Act is less central here than data governance: a normal marketing image is generally not a high-risk AI system, but transparency obligations can still matter when synthetic or manipulated content is used in a context where people could mistake it for authentic material.
European-language support is another open question. Microsoft has not yet published a detailed language matrix for MAI-Image-2.6. Text rendering improvements are welcome, but English lettering quality does not automatically guarantee reliable Czech, German, French or accented European text in packaging and advertising.
Second place is the beginning, not the conclusion
MAI-Image-2.6 has achieved the difficult part: it entered a crowded leaderboard close to the front and substantially improved on its predecessor. The 1,336 score makes Microsoft a serious image-generation competitor rather than a company merely adding another model to a catalogue.
The next test will be practical. Microsoft needs to publish the price, supported resolutions, rate limits, commercial-use terms, regional availability and Foundry deployment details. Until then, GPT Image 2 and Gemini 2.5 Flash Image remain easier to budget. MAI-Image-2.6 has won the attention battle in Arena; the API price will decide whether it also wins a place in European production systems.
Sources: Microsoft AI announcement, Arena Text-to-Image leaderboard, Google Gemini API pricing, and the ECB reference exchange rate, checked August 11, 2026.
