On August 10, 2026, Microsoft AI announced its image generation model "MAI-Image-2.6." In Arena's text-to-image generation ranking—where anonymized models are compared by human evaluators—it took overall 2nd place, behind only OpenAI's GPT Image 2. Compared to its predecessor, MAI-Image-2.5, it gained 79 points in overall score and improved across all 8 evaluated categories. However, as of the announcement, Arena is the only place to try it, and pricing, speed, and detailed availability specifications have not yet been released. The next question is when and under what conditions this improved quality will become available in Microsoft products.
Behind the 2nd-Place Ranking: 3,488 Votes and a "2nd-3rd" Range
In the Arena snapshot updated on August 10, 2026 at 13:00 (UTC), MAI-Image-2.6-preview scored 1335.7, ranking 2nd among 77 models. The top-ranked GPT Image 2 (Medium) scored 1381.1, a gap of about 45.5 points. Grok Imagine Image 2.0 (Low), in 3rd place, scored 1315.5. Microsoft's model also surpassed Google's Nano Banana 2 and Meta's Muse Image.
However, some uncertainty remains in this ranking due to the number of votes. The comparisons involving 2.6 totaled 3,488 votes, considerably fewer than GPT Image 2's 70,065 votes. The score interval is 1324.8–1346.6, and the rank spread (the range of possible rankings calculated from the confidence interval) is 2nd-3rd place. While Arena's raw rank currently gives the best estimate as 2nd place, it cannot yet be said that the statistical difference from 3rd place has been firmly established.
In Arena, users compare two images generated from the same prompt with model names hidden, and vote for their preferred image. Because the accumulated head-to-head results are scored using the Bradley-Terry method, this differs in nature from benchmarks that automatically score fixed problems. While it effectively captures how real users choose between finished images, it is not a direct measure of copyright risk, safety, generation speed, or API pricing.
Improvements Across All 8 Categories, with 3D Rising by 114 Points
In the comparison presented in Microsoft's announcement article, 2.6 outperformed 2.5 in every category. The largest gains were seen in 3D modeling at 114 points, portraits at 105 points, and text rendering at 91 points. Arena's human preference scores support Microsoft's claim of having improved text, three-dimensional structure, and character representation all together.
| Arena Category | MAI-Image-2.6 | MAI-Image-2.5 | Gain |
|---|---|---|---|
| Overall | 1335 | 1256 | +79 |
| 3D Modeling | 1362 | 1248 | +114 |
| Cartoon | 1355 | 1270 | +85 |
| Realistic/Cinematic | 1332 | 1252 | +80 |
| Art | 1345 | 1278 | +67 |
| Portrait | 1364 | 1259 | +105 |
| Text Rendering | 1374 | 1283 | +91 |
| Commercial Design | 1344 | 1265 | +79 |
The gains are not concentrated in a single capability. Microsoft explains that, in addition to enhanced text rendering and improvements to portraits and 3D images, output quality has also improved for product, brand, and cinematic use cases. When 2.5 was announced in June 2026, it was marketed for high-quality generation and image editing. The change in 2.6 lies in significantly boosting the ability to generate an initial image from text, while maintaining that production-oriented focus.
Note that the "2nd place" that 2.5 achieved in June was in Arena's image editing ranking. This time, it is 2nd place in text-to-image generation—the same rank, but a different evaluation task. Since comparative figures for image editing have not been published for 2.6, it cannot be determined whether the precise editing capability demonstrated by 2.5 has improved by the same margin.
Multi-Reference Images Announced, Pricing and Speed Still Undisclosed
MAI-Image-2.6 also comes with features not reflected in the Arena score. Microsoft mentioned the ability to handle multiple reference images, grounding to strengthen connections to information, and controls for adjusting reasoning, output format, and resolution. The announcement article states that details will be "shared soon," and has not disclosed the number of input images supported, compatible resolutions, or which products will offer this capability.
A model card that developers could use for comparison is also not yet available. Without an overview of training data, known limitations, or safety evaluations disclosed, suitability for business use cannot be determined from Arena preference alone. For the previous generation, 2.5, Microsoft explained that while it includes input and output filters, it may still exhibit biases originating from training data and could generate plausible but incorrect visual information.
Pricing also remains an open question. MAI-Image-2.5 is offered on Microsoft Foundry at $5 per million tokens for text input, $8 for image input, and $47 for image output. The faster, lower-cost version, 2.5-Flash, is priced at $1.75, $1.75, and $19.50 respectively. Whether 2.6 will inherit the same pricing tier, or whether pricing and processing time will change alongside the quality improvements, has not been disclosed.
From Arena-First to Foundry Rollout
Availability will proceed in three stages. As of the August 10 announcement, text-to-image generation is available via Arena's Direct Mode. It is planned to roll out to MAI Playground later in the same week, and to Microsoft Foundry and other products "soon." Microsoft Learn's Foundry documentation currently lists 2.5-Pro, 2.5-Flash, 2.5, 2e, and 2 as supported models, with 2.6 not yet included.
This time lag separates ranking from procurement decisions. The fact that it reached 2nd place in Arena shows that Microsoft has pushed its in-house-developed image model into the top tier. On the other hand, for enterprises to incorporate it into production workflows, they need to confirm the regions where the API is available and the output resolutions. Rate limits and pricing are also indispensable. Once these are disclosed with the official Foundry release, and as comparisons accumulate beyond the current 3,488 votes, it will become possible to translate 2.6's quality improvements into cost-effectiveness for production environments.