--- title: "Microsoft's MAI-Image-2.6-Flash Edits Images 2.8x Faster Than GPT Image 2" subtitle: "Microsoft's new Flash image model climbs to #3 on the Artificial Analysis editing leaderboard while keeping the same low price as its predecessor." url: "https://alphasignal.ai/news/microsoft-s-mai-image-2-6-flash-edits-images-2-8x-faster-than-gpt-image-2" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Artificial Analysis" company_url: "https://alphasignal.ai/company/artificialanlys" published: "2026-09-04T16:16:03.000Z" updated: "2026-09-21T21:44:20.843Z" type: "News" topics: - "Image" - "Benchmarks" subtopics: - "Image Editing" - "Image Generation" upvotes: 352 read_time: "4 min" source_url: "https://artificialanalysis.ai/image/leaderboard/text-to-image" comments: 0 --- # Microsoft's MAI-Image-2.6-Flash Edits Images 2.8x Faster Than GPT Image 2 > Microsoft's new Flash image model climbs to #3 on the Artificial Analysis editing leaderboard while keeping the same low price as its predecessor. Microsoft has released [MAI-Image-2.6-Flash](https://microsoft.ai/news/pushing-the-quality-cost-frontier-with-mai-image-2-6/) in public preview on Microsoft Foundry alongside the flagship MAI-Image-2.6. At publication, Flash ranks No. 3 for image editing in the Artificial Analysis Image Arena, behind MAI-Image-2.6 and OpenAI’s GPT Image 2 (high), and narrowly ahead of Google’s Nano Banana 2. It ranks No. 8 for text-to-image generation, one place above Nano Banana Pro. Against MAI-Image-2.5-Flash at the same listed price, the new model gains 69 Elo points in text-to-image generation and 34 points in image editing. Elo scores summarize pairwise human preferences, with higher scores indicating that evaluators chose a model’s outputs more often. Image Arena rankings update as votes arrive, so positions can change. ## A faster diffusion stack Microsoft describes both 2.6 models as diffusion systems that support image generation and controlled image editing. They turn random noise into an image aligned with the prompt. Their flow-matching training teaches a smoother path between noise and the target image, which can reduce the number of sampling steps needed to produce an output. The shared 2.6 feature set includes: - Up to five reference images per request for maintaining consistency across products, characters, people and brand elements. - Bing Search grounding for incorporating current information beyond the model’s training data. - Dynamic aspect ratios and output at up to 1.5K resolution. - Targeted object changes, reframing, layout adjustments, text updates and artifact cleanup while preserving surrounding details. ## Prices favor high-volume workloads Microsoft bills both models by token rather than by image. Flash costs less across text input, image input and image output: Microsoft Foundry list pricing | Token type | MAI-Image-2.6-Flash | MAI-Image-2.6 | | --- | --- | --- | | Text input | $1.75 per 1M tokens | $5 per 1M tokens | | Image input | $2.50 per 1M tokens | $8 per 1M tokens | | Image output | $19 per 1M tokens | $38 per 1M tokens | Artificial Analysis estimates a normalized Flash cost of roughly $19.50 per 1,000 output images. Actual charges depend on output-token consumption, including image dimensions. On its editing quality-versus-price chart, Flash sits on the Pareto frontier between Meta’s Muse Image and Microsoft’s flagship. In this context, the frontier contains models for which no tested alternative offers both a lower estimated price and a higher editing score. Both variants are available through Microsoft Foundry in public preview. OpenRouter also lists them through the Azure provider, allowing applications with an existing OpenRouter integration to use the models without adding a separate Foundry integration. Teams preparing production deployments should confirm regional availability, quotas and preview support terms. ## Editing drives the benchmark gains Artificial Analysis reports Flash’s strongest results in three editing categories: - **Reasoning-based edits:** applying changes implied by the instruction, even when every step is unspecified. - **Composition and framing:** handling zoom-outs, reframing and layout rearrangements. - **Identity preservation:** changing faces or characters while keeping the subject recognizable. Across industry use cases, Flash records the highest score in Architecture and Real Estate. It finishes within six Elo points of the leader in Animation and Gaming, Frontier Prompts, and Live-Action Film. Compared with MAI-Image-2.5-Flash, it narrows the gap to the category leader in nine of ten evaluated use cases. ## Restoration remains the weak spot Enhancement and Restoration is the sole editing action in which Flash failed to narrow its deficit to the category leader compared with its predecessor. Developers building restoration-heavy workflows should test representative images against the flagship and competing models before selecting a default. Microsoft reports that Flash generates images 2.8 times faster than GPT Image 2 Medium while carrying a lower estimated per-image cost. Its lower overall benchmark rank reflects the intended tradeoff: less peak fidelity than the flagship in exchange for greater throughput and lower token charges. ## Choose the model by workload Microsoft positions the two variants for different stages of an image pipeline: | Model | Best suited to | Primary tradeoff | | --- | --- | --- | | MAI-Image-2.6 | Final campaign assets, complex edits, commercial design and high-fidelity text rendering | Higher token cost and latency | | MAI-Image-2.6-Flash | Interactive interfaces, rapid iteration, personalization and high-volume pipelines | Lower peak benchmark quality and weaker restoration performance | ## The production tradeoff shifts At listed rates, Flash cuts the flagship’s image-output token price in half while approaching leading editing quality. For catalog generation, personalization and in-product editors, its lower unit cost and latency can outweigh the remaining Elo gap. Workflows that prioritize final-pass fidelity, complex edits or text rendering still favor MAI-Image-2.6. Microsoft has already deployed its image models across consumer and productivity products. The company previously said Bing Image Creator used MAI-Image-2.5 by default for fully in-house generation, PowerPoint used it for image-to-image work, and key OneDrive editing scenarios had adopted it as their default. The 2.6 release exposes that internal model strategy to developers through hosted APIs and third-party routing. ## Sources - [X post by @artificialanlys](https://x.com/artificialanlys/status/2095908763563680105) - [Pushing the quality-cost frontier with MAI-Image-2.6](https://microsoft.ai/news/pushing-the-quality-cost-frontier-with-mai-image-2-6/) - [MAI-Image-2.6 and MAI-Image-2.6-Flash: Quality and speed at production scale](https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/mai-image-2-6-and-mai-image-2-6-flash-quality-and-speed-at-production-scale/4550970) - [MAI-Image-2.6-Flash | Model Catalog | Microsoft Foundry](https://ai.azure.com/catalog/models/MAI-Image-2.6-Flash?publisher=microsoft) - [Artificial Analysis leaderboard tweet](https://x.com/ArtificialAnlys/status/2095908763563680105) - [Microsoft unveils MAI-Image-2.6-Flash for faster and cheaper image generation](https://www.neowin.net/news/microsoft-unveils-mai-image-26-flash-for-faster-and-cheaper-image-generation/)