Microsoft first-party image-generation/editing family available in Microsoft Foundry public preview, with MAI-Image-2.6 for higher-quality workloads and MAI-Image-2.6-Flash for lower-cost/faster inference.
MAI-Image-2.6 is Microsoft AI's current high-quality image model family, paired with MAI-Image-2.6-Flash for lower-latency and lower-cost workloads. Both are available in Microsoft Foundry as public-preview Global Standard deployments for text-to-image generation and image-to-image editing.1
Microsoft documents multi-reference editing with up to five reference images for the standard model, aimed at preserving products, characters, brand marks and other locked creative elements through revisions. The family also adds stronger format/resolution control and web-grounded creation compared with the prior generation.2
The standard 2.6 model is positioned for final assets and precision; Flash is intended for interactive iteration, personalization and high-volume production where throughput matters more. This is a useful two-model production pattern, but buyers should treat Microsoft's speed and efficiency comparisons as vendor claims rather than independent measurements.2
Microsoft Foundry now publishes exact usage pricing. MAI-Image-2.6 starts at $5 per 1M text-input tokens, $8 per 1M image-input tokens and $38 per 1M image-output tokens. MAI-Image-2.6-Flash starts at $1.75, $2.50 and $19 respectively.2
Both are public-preview Foundry models rather than open-weight releases. Deployment is through Global Standard endpoints, with region availability and Azure billing rules applying.1 The Flash price materially improves the family's suitability for iterative or scaled generation, while standard 2.6 targets final-quality output.
Independent blind-vote evidence is unusually strong for such a recent model. In the current Artificial Analysis Text to Image Arena snapshot, MAI-Image-2.6 ranks second with Elo 1149 across 6,351 samples; the Editing Arena lists it at Elo 1122 across 11,173 samples. Flash appears lower in text-to-image but strongly in editing at a much lower representative API price.34
Those numbers are useful evidence of average preference under the benchmark protocol, not a guarantee for a specific commercial workflow. The current AiToolMap panel explicitly configures Artificial Analysis as editorial evidence rather than a numeric rating input, so its Elo values are not blended into the public score.
A separate small practical comparison, surfaced through a Hugging Face-hosted dataset, found a more mixed picture: MAI-Image-2.6 Preview produced that tester's favorite product advertisement but reportedly failed to return one hospital-screen scene after roughly ten attempts. The test is narrow and provider/context-dependent, so it is qualitative evidence of possible reliability variance rather than a failure-rate estimate.5
The 50-source canonical panel found no other current exact-product evidence that qualifies for numeric transfer. Seven sources were inaccessible through the lawful crawler path and were not bypassed. No AiToolMap hands-on testing is claimed.
MAI-Image-2.6 fits developers, design systems and marketing/product teams already comfortable with Azure/Microsoft Foundry and needing strong image generation plus editing in an API workflow. Multi-reference support makes it particularly relevant to branded creative, product imagery and character/asset consistency.2
Flash is the more natural choice for interactive applications, personalization, rapid concept iteration and production volumes where latency and token price dominate; standard 2.6 is the better starting point when maximum output quality and edit fidelity are the priority.2
current top-tier blind-preference performance, both generation and editing support, multi-reference control, web grounding, a clearly differentiated Flash variant and transparent Foundry token pricing.342 Microsoft's enterprise deployment channel also makes the family easier to integrate into existing Azure procurement and operations than consumer-only image tools.
public-preview maturity, limited independent longitudinal reliability evidence and the need to validate behavior on the buyer's own brand/text/reference cases. The small practical test's task failure is a reminder that high Arena position can coexist with brittle edge cases.5 AiToolMap therefore scores the family highly on capability/value but keeps confidence LOW until more exact-product production evidence accumulates.
Products matched by shared tasks, audience, workflow, product type and capabilities.