← Back to feed

Microsoft released two in-house AI models to public preview this week: MAI-Image-2.5-Pro, its highest-fidelity image model, and MAI-Voice-2-Flash, a speech model built for high-volume enterprise use. The headline claim across both: up to 89% lower cost than comparable OpenAI offerings for the same tasks.

MAI-Voice-2-Flash targets the workloads where audio inference spend is highest: call center transcription, real-time voice pipelines, and batch audio processing at enterprise scale. At 89% below OpenAI’s equivalent pricing, it would undercut the Realtime API for any deployment where strict real-time latency is not required.

The Numbers So Far

MAI-Image-2.5-Pro entered Arena’s text-to-image leaderboard at Elo 1,254, ranking third — the first Microsoft model to reach the top tier on that benchmark. No Arena leaderboard entry exists yet for MAI-Voice-2-Flash; the model went into public preview without a published evaluation ranking.

Microsoft has not released per-token pricing for MAI-Voice-2-Flash. The 89% cost reduction figure applies to specific task comparisons; full cost benchmarks across all workload types are not yet published. OpenAI’s Realtime API is currently priced at $0.06 per minute of audio input. Whether MAI-Voice-2-Flash pricing holds at scale and across varied audio conditions depends on the full pricing table, which is pending.

A Structural Contradiction

Microsoft holds a reported $13B+ investment in OpenAI and built the Azure infrastructure that powers GPT model serving. Azure OpenAI Service has been Microsoft’s primary vehicle for distributing OpenAI products to enterprise customers since 2023.

MAI-Voice-2-Flash changes that arrangement. A model that undercuts OpenAI pricing by 89% — hosted on the same Azure platform, sold to the same enterprise buyers — is a competing product, not a complementary one. Microsoft frames the MAI family as tools to reduce enterprise AI costs. The strategic reading is that Azure’s AI revenue is no longer dependent on OpenAI’s pricing being the floor.

This is not the first time the two companies have been in tension. The Apple-OpenAI partnership fracture and OpenAI’s push toward direct enterprise deals have both created friction with Microsoft’s distribution model. MAI-Voice-2-Flash adds in-house product competition to that list.

What MAI-Voice-2-Flash Is Built For

Speech AI at enterprise scale involves two distinct cost profiles. Real-time voice — live call transcription, voice assistants, meeting tools — requires low latency and typically justifies higher per-minute pricing. Batch audio — recorded call analytics, compliance transcription, audio archiving — is latency-tolerant and price-sensitive at volume.

MAI-Voice-2-Flash, built for high-volume enterprise use, is likely optimized for the batch and moderate-latency segment. The “Flash” naming convention across Microsoft’s MAI line follows the industry pattern (Gemini Flash, Claude Haiku) of tagging faster, cheaper variants designed for throughput-first workloads.

Public Preview Status

Both MAI models are currently in public preview on Azure AI. General availability timing has not been announced. Microsoft’s preview-to-GA timeline for Azure AI models has historically run two to six months. Enterprise customers evaluating alternatives to OpenAI’s Realtime API or DALL-E pricing would be the natural early adopters.

Key Numbers

• Models released: MAI-Voice-2-Flash (speech), MAI-Image-2.5-Pro (image)
• Status: Public preview
• Cost claim: Up to 89% cheaper than OpenAI equivalents
• MAI-Image-2.5-Pro Arena ranking: #3, Elo 1,254 (text-to-image)
• MAI-Voice-2-Flash Arena ranking: None published yet
• Per-token pricing: Not yet published for MAI-Voice-2-Flash