Microsoft launches MAI-Image-2.6 and MAI-Transcribe 2 on OpenRouter

Microsoft's most powerful image-generation model has landed on OpenRouter in version 2.6, bringing multi-image editing and real-time web grounding. The model supports dynamic aspect ratios and outputs up to 1.5K pixels on the long side, giving production workflows more flexibility than earlier versions that locked users to fixed resolutions.

A flash variant, MAI-Image-2.6-Flash, ships alongside the full model. Built for latency-sensitive production environments, it promises higher throughput without dropping the feature set — multi-image editing, web grounding, and flexible aspect ratios remain available. Microsoft AI has not published comparative benchmarks against competitors.

On the audio side, MAI-Transcribe 2 takes the top spot on the multilingual FLEURS benchmark. It covers 60 languages with automatic language detection and code-switching support — speech that mixes languages within a single sentence. The model targets subtitles, conversation transcription, accessibility, and voice applications, and Microsoft says it is faster than its predecessor, MAI-Transcribe-1.5, on long-form audio files.

The new model adds speaker diarization to separate talkers, word-level timestamps, and keyword biasing for domain-specific terminology such as product names or medical jargon. Developers can choose between verbatim and clean transcription styles, all controlled through dedicated API parameters.

On OpenRouter, features are activated via a response_format of verbose_json to receive segment timestamps, with timestamp_granularities: ["word"] for word-level granularity. Diarization is enabled through provider.options.azure.diarization.enabled, keyword lists pass through provider.options.azure.phraseList.phrases, and transcription style selection uses provider.options.azure.enhancedMode.modelOptions.transcribeStyle — all documented in the platform's speech-to-text guide.