MAI-Image-2.6 is a text-to-image and image-editing model built in-house by Microsoft AI, the division led by Mustafa Suleyman. Microsoft announced it on August 10, 2026, saying it had debuted at No. 2 on the Arena text-to-image leaderboard with a gain of 79 Elo points over its predecessor, MAI-Image-2.5. [1] On September 4, 2026, Microsoft opened a public preview of the model in Microsoft Foundry and added a second variant, MAI-Image-2.6-Flash, which the company describes as an optimized version of the same model built for latency-sensitive, high-throughput workloads. [2][4]

The model card published on September 4 describes MAI-Image-2.6 as a diffusion-based generator trained with a flow-matching loss, with 20 billion non-embedding parameters, a 32K-token context and a maximum output of 2,359,296 pixels (the equivalent of 1536 by 1536). Training ran from April 17 to July 22, 2026. [4] Both variants support editing from multiple reference images, web grounding, and dynamic aspect ratios. [2]

Microsoft's headline claims for the September release are that MAI-Image-2.6 ranks No. 2 for both text-to-image and image editing on Arena and No. 2 for text-to-image and No. 1 for image editing on Artificial Analysis, both footnoted "as of Sep 4, 2026", and that MAI-Image-2.6-Flash "is able to generate images 2.8x faster than GPT-Image-2-Medium while delivering 72% greater efficiency". [2] Suleyman's own post the same day rounded the speed figure to "2x faster than GPT-Image-2". [6] Independent leaderboard snapshots taken on September 5, 2026 broadly agree with the rank claims, although the two leaderboards score models on different scales and neither shows Microsoft's model ahead of GPT Image 2 in text-to-image. [7][8][9][10]

Key facts

Attribute	Detail
Developer	Microsoft AI (Microsoft Corporation; EU representative Microsoft Ireland Operations Limited) [4][5]
Variants	MAI-Image-2.6; MAI-Image-2.6-Flash ("an optimized version of MAI-Image.2.6", per the model card) [4]
Announced	August 10, 2026 (Arena debut); model card lists a release date of August 14, 2026 for 2.6 and September 4, 2026 for 2.6-Flash [1][4]
Foundry public preview	September 4, 2026; model version `2026-07-31`, Global Standard deployment [2][11]
Predecessor	MAI-Image-2.5 (June 2, 2026) [1][21]
Architecture	Diffusion-based text-to-image and image-to-image model trained with a flow-matching objective (company-reported) [4]
Parameters	2 x 10^10 (20B) non-embedding parameters (company-reported) [4]
Context length	32K tokens [4][12]
Inputs and outputs	Text and image (JPEG or PNG) in; one PNG image out [12]
Max resolution	"Up to 1.5K" per the launch post; model card caps total pixels at 2,359,296 (1536 x 1536 equivalent); Foundry documentation lists 1,048,576 total pixels (1024 x 1024 equivalent) as of September 5, 2026 [2][4][12]
Training window	April 17 to July 22, 2026; data collected as late as May 2026 [4][5]
Availability	Microsoft Foundry (public preview), MAI Playground, Arena [1][2][3]
Leaderboards (company claim)	Arena: No. 2 text-to-image and image editing; Artificial Analysis: No. 2 text-to-image, No. 1 image editing, as of September 4, 2026 [2]

Announcement and rollout

The first public sign of MAI-Image-2.6 was a Microsoft AI post on August 10, 2026 titled "MAI-Image-2.6 launches at No. 2 on Arena ahead of Google, Meta and xAI". According to the post, the model ranked second on Arena's text-to-image leaderboard, improved by 79 Elo over MAI-Image-2.5 overall "with gains across every Arena text-to-image category", and improved text rendering alone by 91 Elo. Microsoft wrote that the result "firmly establishes MAI-Image ahead of leading models from Meta, Google and xAI". [1] The post listed three areas of improvement: stronger text rendering; better portraits and 3D imagery; and "more polished commercial and photorealistic outputs, with stronger results across product, branding and cinematic use cases". It also teased features that were held back at the time, "from working across multiple references and richer grounding to greater control over reasoning, format and resolution". On August 10 the model was available only for text-to-image on Arena, with MAI Playground promised "later this week" and Foundry "soon". [1]

Microsoft's model card and its EU data summary both give August 14, 2026 as the model's release date, four days after the Arena announcement. [4][5] The Artificial Analysis leaderboard records a release date of August 10 for the model and added it to its board on August 12. [7] The data summary lists September 4, 2026 as the "date of placement of the model on the Union market", the same day as the Foundry public preview. [5]

On August 25, 2026, Artificial Analysis reported on X that a model then labelled MAI-Image-2.6-Preview had taken the No. 1 position on its Image Editing leaderboard and No. 2 in text-to-image; the post was reproduced by 24/7 Wall St. [14] Microsoft's Foundry catalog entry for the model, which still carries copy from the August period, says the model "ranked No. 2 on the Arena text-to-image leaderboard" and had "achieved No. 3 on Arena's image editing leaderboard, ahead of Google's Nano Banana family and Meta's Muse Image". [13]

The second announcement came on September 4, 2026, under the title "Pushing the quality-cost frontier with MAI-Image-2.6". Microsoft called MAI-Image-2.6 "our strongest image model yet", brought it to developers in Microsoft Foundry as a public preview, and introduced MAI-Image-2.6-Flash. The August 10 post was updated the same day with a banner noting that both models were now in public preview on Foundry. [1][2] Microsoft Learn documentation for the MAI image models was last updated on September 4 and lists both 2.6 models with model version 2026-07-31. [11]

What changed in 2.6

Microsoft's September post describes three capability additions shared by both variants. Multi-reference editing "brings together people, products, styles, and scenes from different images". Web grounding "pulls info from across the web to create rich visuals informed by relevant, up-to-date information". Higher resolutions and dynamic aspect ratios let the model choose "the optimal format for compositions with support for up to 1.5K resolution". [2] The model page repeats the web-grounding pitch as "grounds creativity in current context by using the latest web content and documents", and illustrates it with a prompt for a wide-angle photograph of Madrid's Metropolis Building "as it stands today". [3]

The model card gives the technical framing. MAI-Image-2.6 "progressively transforms random noise into a coherent image aligned with a given text prompt, leveraging a flow-matching loss to learn a continuous transformation between the noise distribution and the data distribution". Microsoft says the model "was rebuilt from the ground up for multimodal editing workflows, with stronger visual context understanding and coherence", and that it "reasons across objects, scene structure, lighting, scale, and spatial positioning to produce consistent edits". Supported edit types listed on the card are object removal, replacement, attribute changes, inpainting, text updates and artifact cleanup such as removing motion blur, "without destabilizing composition or layout". [4]

Specification (model card, September 4, 2026)	Value
Model architecture	Diffusion-based generative architecture for text-to-image synthesis and image-to-image editing
Parameters	2 x 10^10 (20B) non-embedding parameters
Inputs	Text input; image input (for editing workflows)
Context length	32K tokens
Outputs	Image output; maximum total pixel count 2,359,296 (equivalent to 1536 x 1536), either dimension may exceed 1536 within that total
Training dates	April 17, 2026 to July 22, 2026
Release date	Aug 14, 2026 (2.6), Sep 4 (2.6-Flash)
Release date in the EU	September 4, 2026
License	"Various product and service terms where the model is deployed, such as those for Azure AI Foundry and MAI Playground"
Model dependencies	"MAI-Image-2.6-Flash is an optimized version of MAI-Image.2.6"

Source: Microsoft's model card. [4]

The card's evaluation section says the model "was evaluated by human raters alongside comparable models and across a range of capability areas", with raters choosing a preferred output for topic areas such as "product/branding", "cartoon" and "photorealistic", producing an Elo score. Its safety section says the model was tested "across a range of adversarial safety scenarios, including higher-severity content" with results "broadly consistent with the prior model version", and closes with the note that "this model card will be updated as training completes and evaluation data becomes available". [4]

MAI-Image-2.6-Flash

MAI-Image-2.6-Flash is the throughput-oriented sibling. Microsoft says it "brings comparable quality to latency-sensitive, high-throughput workloads" and "is able to generate images 2.8x faster than GPT-Image-2-Medium while delivering 72% greater efficiency". The post does not publish the test conditions behind either figure, and both are Microsoft's own measurements. [2] The model page describes Flash as "production-ready quality at less than half the price of our flagship model". [3] Artificial Analysis's representative API prices as of September 5, 2026 were $38.9 per 1,000 images for MAI-Image-2.6 and $19.5 per 1,000 for MAI-Image-2.6-Flash, roughly half. [7][8]

Suleyman promoted the Flash model on X on September 4, 2026. The post reads in full:

Our new image model generates images 2x faster than GPT-Image-2, currently the best model in the world.

It's also 72% more efficient in GPU usage, so we can provide it at an incredible price.

This gives it the best price-performance score in the world.

Unbelievable work from the team. So much more to come! Try MAI-Image-2.6-Flash out now! [6]

The two Microsoft statements differ in wording. The blog post says "2.8x faster than GPT-Image-2-Medium", naming a specific quality tier of OpenAI's model; Suleyman's post says "2x faster than GPT-Image-2" without a tier and describes GPT-Image-2 as "currently the best model in the world". The blog says "72% greater efficiency"; the post says "72% more efficient in GPU usage", which is the only place the efficiency figure is tied to GPU usage. The blog's price claim is "the best price-per-Elo performance in the world"; the post's is "the best price-performance score in the world". Microsoft has not published a reconciliation. [2][6]

The post carried a chart titled "Quality vs. Price" with the subtitle "Image Arena Elo vs. representative API price per 1,000 images" and a source line crediting "Artificial Analysis Arena". The chart shades a "most attractive quadrant" and draws a dotted "Pareto line". MAI-Image-2.6 sits at the top of the plotted field at a price of roughly $40 to $50 per 1,000 images, MAI-Image-2.6-Flash a little lower at roughly $25, and both are marked as points on the Pareto line together with Muse Image, the cheapest model plotted at about $10. Nano Banana 2 (Gemini 3.1 Flash Image) is plotted just below MAI-Image-2.6 at a higher price, and GPT Image 2 (high) is plotted far to the right at around $200 with an Elo close to Nano Banana 2's. MAI-Image-2.5, MAI-Image-2.5-Pro and MAI-Image-2.5-Flash, Seedream 5.0 Pro, a point labelled Nano Banana 3 (Gemini Pro Image), GPT Image 1.5 (high), Qwen-Image-3.0-Pro, grok-imagine-image-quality, Reve 2.1 and Luma UNI 1 Max are plotted lower. The chart's vertical axis runs from 1260 to 1330, which does not match the scale of the Artificial Analysis leaderboards as fetched on September 5 (where MAI-Image-2.6 scores 1149 in text-to-image and 1126 in editing), and its ordering, with MAI-Image-2.6 above GPT Image 2 (high), matches the editing board rather than the text-to-image board. Exact values should be read from the leaderboards rather than from the image. [6][7][8]

Artificial Analysis posted on September 4 that MAI-Image-2.6-Flash "takes #3 on the Artificial Analysis Image Editing Leaderboard, a significant jump over the previous generation's Flash variant", and that Microsoft now had two models "on the Pareto frontier for quality vs price". [15] Its editing rank moved between second and fourth during its first two days on the board, as the Elo values shifted by several points within hours. [8]

Benchmarks and leaderboards

Microsoft's claims are all drawn from the two public human-preference arenas. The company reported the following, in each case as of the date given. [1][2]

Claim (Microsoft)	Leaderboard	Date given
No. 2 text-to-image; +79 Elo over MAI-Image-2.5 overall; +91 Elo in text rendering	Arena	August 10, 2026
No. 2 text-to-image and No. 2 image editing	Arena	September 4, 2026
No. 2 text-to-image and No. 1 image editing	Artificial Analysis	September 4, 2026
Flash: 2.8x faster than GPT-Image-2-Medium, 72% greater efficiency	Microsoft measurement	September 4, 2026
"Best price-per-Elo performance in the world"	Microsoft, citing Arena Elo and its own pricing	September 4, 2026

The Arena text-to-image leaderboard fetched on September 5, 2026 listed gpt-image-2 (medium) first with a rating of 1381.8 on 77,830 votes, and mai-image-2.6 next with 1332.2 on 10,086 votes, in a rank band of 2 to 3 shared with grok-imagine-image-2.0 (low) at 1315.3. Reve 2.1 (1301.2), Muse Image (1279.2) and Reve 2.0 (1270.3) followed. Earlier MAI models sat further down: mai-image-2.5 at 1254.1 (rank band 7 to 11), mai-image-2 at 1182.6 (16 to 20) and mai-image-1 at 1093.3 (51 to 53). [9] Arena's image-editing leaderboard on the same day had gpt-image-2 (medium) first at 1461.0 on 228,827 votes, then mai-image-2.6 at 1438.9 on 5,086 votes and grok-imagine-image-2.0 (low) at 1438.8, both in a rank band of 2 to 3, followed by Muse Image (1403.5) and mai-image-2.5 (1400.2). No MAI-Image-2.6-Flash entry appeared on either Arena board as of that fetch. [10]

The Artificial Analysis boards use a different Elo scale. Its overall text-to-image board on September 5, 2026 read as follows; the tables below are a same-day snapshot, and Artificial Analysis Elos and ranks shifted by several points within hours on September 5. [7]

Rank	Model	Elo	Representative price per 1,000 images
1	GPT Image 2 (high)	1177	$211.0
2	MAI-Image-2.6	1149	$38.9
3	Reve 2.1	1127	$200.0
4	Nano Banana 2 (Gemini 3.1 Flash Image)	1121	$67.0
5	Muse Image	1115	$10.0
6	GPT Image 1.5 (high)	1113	$133.0
7	MAI-Image-2.5	1112	$48.1
8	MAI-Image-2.6-Flash	1100	$19.5
9	Nano Banana Pro (Gemini 3 Pro Image)	1099	$108.5
10	MAI-Image-2.5-Pro	1097	$134.0

Source: Artificial Analysis text-to-image leaderboard, September 5, 2026; the Flash price is taken from the same site's editing board. Older MAI models ranked 20th (MAI-Image-2.5-Flash, 1031), 31st (MAI-Image-2, 1011), 49th (MAI-Image-2-Efficient, 984) and 114th (MAI Image 1, 863). [7][8]

The Artificial Analysis image-editing board on September 5, 2026 placed MAI-Image-2.6 first. [8]

Rank	Model	Elo	Representative price per 1,000 images
1	MAI-Image-2.6	1126	$38.9
2	GPT Image 2 (high)	1120	$211.0
3	Nano Banana 2 (Gemini 3.1 Flash Image)	1111	$67.0
4	MAI-Image-2.6-Flash	1111	$19.5
5	Muse Image	1110	$10.0
6	MAI-Image-2.5	1100	$48.1
7	Seedream 5.0 Pro	1099	$90.0
8	MAI-Image-2.5-Pro	1098	$134.0

Source: Artificial Analysis image-editing leaderboard, September 5, 2026. [8]

Two caveats apply to all of these figures. Arena and Artificial Analysis run separate voting pools with different prompt sets, quality settings and Elo baselines, so a rank on one is not comparable to a rank on the other, and the numbers move as votes accumulate; Microsoft's own posts footnote each rank with a date for that reason. [2] Second, Microsoft's comparison target differs by leaderboard: Arena lists GPT Image 2 at its "medium" setting while Artificial Analysis lists the "high" setting, and Microsoft's speed claim names the medium tier. [2][9][7]

Pricing and availability

MAI-Image-2.6 and MAI-Image-2.6-Flash are sold as "Foundry Models sold directly by Azure". Microsoft Learn lists both as preview models with model version 2026-07-31 and Global Standard deployment, exposing text-to-image generation and image-to-image edits through a dedicated MAI images API (/mai/v1/images/generations and an edits endpoint) with width, height and prompt parameters. The documented constraints as of September 5, 2026 are text or image (JPEG or PNG) input, one PNG image output, a 32,000-token context, English only, no tool calling, and a minimum of 768 by 768 pixels with a maximum total pixel count of 1,048,576. That last figure is the same one the documentation lists for the MAI-Image-2.5 family and is lower than the 2,359,296-pixel cap in the model card and the "up to 1.5K" wording of the launch post; it is not clear whether the documentation had been updated for 2.6 at the time of writing. [2][4][11][12]

Microsoft has not published per-token prices for the 2.6 models in its announcement posts, which state only that Flash costs "less than half the price of our flagship model" and direct developers to Foundry for pricing. [2][3] A query of the Azure Retail Prices API on September 5, 2026 returned meters for MAI-Image-2, MAI-Image-2e, MAI-Image-2.5, MAI-Image-2.5-Flash and MAI-Image-2.5-Pro, but none yet for the 2.6 models. [24] For comparison, Microsoft's published 2.5-generation prices were $5 per 1M text input tokens, $8 per 1M image input tokens and $47 per 1M image output tokens for MAI-Image-2.5; $1.75, $1.75 and $19.50 for MAI-Image-2.5-Flash; and $5, $8 and $106 for MAI-Image-2.5-Pro. [21][22]

Both models can also be used in the MAI Playground, Microsoft AI's consumer-facing site for trying MAI models, and MAI-Image-2.6 remains available for blind voting on Arena. [1][2][3] The model card lists the distribution channels as "Azure AI Foundry - API access for developers (private preview, expanding to public preview)" and "MAI Playground - publicly available site for users to interact with MAI models". [4]

Training data and safety

Microsoft published a public data summary for the two models in the format required of general-purpose AI model providers under the EU AI Act; the document is dated August 14, 2026 and was last updated on September 4, 2026, and Microsoft states that it is a signatory to the Code of Practice for general-purpose AI models. [5] According to the summary, the text training corpus falls in the 1 billion to 10 trillion token range and consists mainly of "safety-filtered, supervised fine-tuned (SFT)/synthetic image caption dataset collections", and the image corpus exceeds 1 billion images drawn from "safety-filtered, acquired, open source, or publicly available image dataset collections curated for quality". Text data is in English. Data was collected as late as May 2026, the training set was first used in April 2026, and the model is not updated with new data after deployment. [5]

The summary names Wikipedia as a large publicly available dataset, says Microsoft "leveraged data acquisition agreements" for licensed image data and bought private image datasets from third parties under confidentiality terms, and describes web crawling with Bingbot between February 2024 and May 2026 that respects robots.txt and other reservation-of-rights signals, excludes paywalled content and domains on the USTR Notorious Markets list, and filters crawled data for deduplication, watermarks, quality and aesthetics. The top 10 percent of crawled domains is described as a mix of image hosts, blogging and social platforms, e-commerce sites and region-specific portals across country codes including .jp, .de, .ru, .cn, .fr, .br, .it, .uk, .pl and .kr. Synthetic data was used to fill domains where real data is scarce: vision-language models produced captions (Microsoft gives the example of GPT-4o captioning approximately 1,000 images on self-hosted infrastructure), and an open-source 3D creation suite was used to render images containing words such as billboards to improve text rendering. User data from Microsoft products was not used. [5]

On safety, the model card says Microsoft "took a defense-in-depth approach, applying mitigations to the data during model development, and deploying the model with additional safety mitigations", with prompt and output filtering to block harmful, abusive or policy-violating content. Listed out-of-scope uses are content intended to deceive, mislead or impersonate real individuals, content that violates applicable laws or Microsoft's terms, and harmful, abusive or policy-violating content. [4]

MAI-Image lineage

MAI-Image-2.6 is the sixth named release in a line that began in October 2025. Every model in the family was developed in-house by Microsoft AI, and each launch post has led with a human-preference leaderboard result. [1][17]

Model	Announced	Availability at launch	Microsoft's headline claim
MAI-Image-1	October 13, 2025	Arena voting; Bing Image Creator and Copilot Audio Expressions from November 4, 2025	"Our first image generation model developed entirely in-house, debuting in the top 10 text-to-image models on LMArena" [17]
MAI-Image-2	March 19, 2026 (MAI Playground); Microsoft Foundry from April 2, 2026	MAI Playground; Copilot and Bing Image Creator rollout; Foundry at $5 per 1M text input tokens and $33 per 1M image output tokens	"Pushing MAI into the top three text-to-image labs in the world on the Arena.ai leaderboard"; "at least 2x faster generation times" on Foundry and Copilot [18][19]
MAI-Image-2-Efficient	April 14, 2026	Microsoft Foundry and MAI Playground; $5 per 1M text input tokens, $19.50 per 1M image output tokens	"22% faster and 4x more efficient" than MAI-Image-2, "priced nearly 41% lower", and "40% faster on average than other leading text-to-image models" (as tested April 13, 2026) [20]
MAI-Image-2.5 and MAI-Image-2.5-Flash	June 2, 2026	Foundry, MAI Playground, OpenRouter; PowerPoint and OneDrive	No. 3 text-to-image and No. 2 image editing on Arena; +75 points over MAI-Image-2 with text rendering +107; 2.5 at $5/$8/$47 per 1M tokens, Flash at $1.75/$1.75/$19.50 [21]
MAI-Image-2.5-Pro	July 23, 2026	Foundry public preview; $5/$8/$106 per 1M tokens	"Our highest-fidelity image model to date"; Bing Image Creator "100% in-house by default" on MAI-Image-2.5 [22]
MAI-Image-2.6	August 10, 2026 (model card release date August 14)	Arena; MAI Playground later that week; Foundry public preview September 4	No. 2 on Arena text-to-image, +79 Elo over 2.5, text rendering +91 Elo [1]
MAI-Image-2.6-Flash	September 4, 2026	Foundry public preview and MAI Playground	"2.8x faster than GPT-Image-2-Medium while delivering 72% greater efficiency"; "less than half the price" of MAI-Image-2.6 [2][3]

MAI-Image-1 was introduced in October 2025 as the first image generator the company had "developed entirely in-house", with Microsoft emphasizing photorealistic lighting and a deliberate effort to avoid "repetitive or generically-stylized outputs"; it entered Bing Image Creator on November 4, 2025 alongside DALL-E 3 and GPT-4o in the model menu. [17] MAI-Image-2 followed in March 2026 with a focus on photorealism, in-image text and detailed scene generation, initially through the MAI Playground and a limited API for customers such as WPP, before reaching Foundry on April 2, 2026. [18][19] Two weeks later Microsoft added MAI-Image-2-Efficient, its first speed-and-cost variant, with a chart of median render times measured on April 13, 2026 against Gemini 3.1 Flash, Gemini 3.1 Flash Image, Gemini 3 Pro Image and GPT-Image-1.5-High. [20]

MAI-Image-2.5 and its Flash variant arrived at Build on June 2, 2026, adding fine-grained edit control and face and identity consistency and going live inside PowerPoint and OneDrive. [21] MAI-Image-2.5-Pro, in July, was presented as the quality-first option; the same post said MAI-Image-2.5 was cutting PowerPoint GPU costs "up to 84% compared with GPT-Image-2" and had raised OneDrive save rates by 26 percent, figures that are Microsoft's own. [22] MAI-Image-2.6 therefore inherits a three-tier pattern (flagship, Flash, and in the 2.5 generation a Pro) that Microsoft describes as giving "every product the right balance of quality, speed, and cost". [22]

Context: Microsoft AI's in-house models

The MAI-Image series is one strand of a broader program under Suleyman, who joined Microsoft in March 2024 and since November 2025 has led the MAI Superintelligence team. Microsoft AI has released in-house text, voice, transcription, coding, reasoning and image models under the MAI prefix, including MAI-1-preview, MAI-Voice-1, MAI-Code-1 and MAI-Thinking-1, and describes them as trained on "clean, traceable, enterprise-grade data, without distillation from third-party models". [22] The company's stated goal is to run its own products on its own models where quality and cost allow, while continuing to use OpenAI and Anthropic models for more demanding tasks. [23]

Image generation is where that substitution is furthest along. Microsoft said in July 2026 that Bing Image Creator was "100% in-house by default" on MAI-Image-2.5, that the same model was in production in PowerPoint for image-to-image work, and that it had become the default for key OneDrive editing scenarios. [22] eWeek, citing a Bloomberg interview, reported on July 27, 2026 that Suleyman said MAI models were deployed in more than half of Microsoft's products, and quoted him: "It's faster, it's cheaper, it's higher quality, it drives better retention". [23] Microsoft's repeated benchmarking of its Flash models against GPT-Image-2, which its own products previously used, reflects that positioning; Suleyman's September 4 post still called GPT-Image-2 "currently the best model in the world". [6]

The 2.6 launch also drew the Artificial Analysis leaderboard into Microsoft's marketing more directly than earlier releases, which had cited Arena alone. Microsoft's Foundry catalog entry for 2.6 still leads with the Arena result, while the September 4 post cites both boards with matching dates. [2][13]

Reception

Coverage of the August announcement was thin and mostly relayed Microsoft's Arena claim. The more substantive independent signal came from Artificial Analysis, whose August 19, August 25 and September 4 posts tracked MAI-Image-2.5-Pro, MAI-Image-2.6-Preview and MAI-Image-2.6-Flash onto its editing leaderboard; 24/7 Wall St reproduced each post for an investor audience, framing the results as "hard evidence that its plan to swap third-party AI out of its own products is working" and, in September, as a sign that Microsoft's "own stack is closing the gap with the OpenAI models it licenses". [14][15] Neowin's September 4 headline and summary relayed Microsoft's 2.8x-faster and 72 percent efficiency figures (the article body was not accessible for this entry). [16]

No third party had published an independent speed or cost measurement of MAI-Image-2.6-Flash against GPT-Image-2-Medium as of September 5, 2026, so the 2.8x and 72 percent figures remain Microsoft's alone. Windows Report's September 5 write-up likewise relayed the company's framing, saying Microsoft "says Flash generates images more than twice as fast as GPT-Image-2-Medium while delivering 72% greater efficiency", and noted that both models accept up to five reference images per request.[25] The one number an outside party has published in the same territory is Artificial Analysis's price-per-1,000-images comparison, which as of September 5 listed the Flash model at $19.5 against $211.0 for GPT Image 2 (high) and $38.9 for MAI-Image-2.6. [7][8]

References

MAI-Image-2.6 launches at No. 2 on Arena ahead of Google, Meta and xAI - Microsoft AI, August 10, 2026 (updated September 4, 2026).
Pushing the quality-cost frontier with MAI-Image-2.6 - Microsoft AI, September 4, 2026.
MAI-Image-2.6 model page - Microsoft AI, accessed September 5, 2026.
MAI-Image-2.6 / MAI-Image-2.6-Flash Model Card (PDF) - Microsoft, September 4, 2026.
Data Summary for MAI-Image-2.6 / 2.6-Flash (PDF) - Microsoft, August 14, 2026 (last updated September 4, 2026).
Post on MAI-Image-2.6-Flash - X (Mustafa Suleyman), September 4, 2026.
Text to Image Leaderboard - Artificial Analysis, accessed September 5, 2026.
Image Editing Leaderboard - Artificial Analysis, accessed September 5, 2026.
Text-to-Image Leaderboard - Arena, accessed September 5, 2026.
Image Editing Leaderboard - Arena, accessed September 5, 2026.
Deploy and use MAI image models in Microsoft Foundry - Microsoft Learn, updated September 4, 2026.
Foundry Models sold directly by Azure - Microsoft Learn, accessed September 5, 2026.
MAI-Image-2.6 - Model Catalog - Microsoft Foundry, accessed September 5, 2026.
Microsoft MAI-Image-2.6-Preview takes the top image editing rank - 24/7 Wall St., August 25, 2026.
Microsoft MAI-Image-2.6-Flash hits No. 3 on the AI image editing leaderboard - 24/7 Wall St., September 4, 2026.
Microsoft unveils MAI-Image-2.6-Flash for faster and cheaper image generation - Neowin, September 4, 2026.
Introducing MAI-Image-1, debuting in the top 10 on LMArena - Microsoft AI, October 13, 2025 (updated November 4, 2025).
Introducing MAI-Image-2: for limitless creativity - Microsoft AI (MSI team), March 19, 2026.
Announcing 3 new world class MAI models, available in Foundry - Microsoft AI (Mustafa Suleyman), April 2, 2026.
MAI-Image-2-Efficient: Flagship Quality, 41% Lower Cost - Microsoft AI (MAI Superintelligence Team), April 14, 2026.
MAI-Image-2.5 launches at No. 2 for image editing on Arena - Microsoft AI (Superintelligence team), June 2, 2026.
Introducing MAI-Image-2.5-Pro and MAI-Voice-2-Flash - Microsoft AI (Superintelligence team), July 23, 2026.
Microsoft's In-House AI Replaces OpenAI Models in PowerPoint, Bing and OneDrive - eWeek (Aminu Abdullahi), July 27, 2026.
Azure Retail Prices API, MAI product meters - Microsoft Azure, queried September 5, 2026.
Microsoft's New MAI-Image-2.6-Flash Is More Than Twice as Fast and Far Cheaper - Windows Report (Milan Stanojevic), September 5, 2026.

Editing: MAI-Image-2.6