Microsoft moved two more in-house models out of preview and into general access on July 23. Both are positioned as substitutes for models the company currently licenses from outside labs to run its own products. Microsoft AI, the internal lab building Microsoft’s proprietary model family, announced the release in a blog post that doubled as a scorecard for where those substitutions have already happened. MAI-Image-2.5-Pro is Microsoft’s highest-fidelity image model to date, built for hero imagery, detailed editing, and precise in-image text rendering. MAI-Voice-2-Flash, a faster and cheaper sibling to the MAI-Voice-2 model already in production, targets high-volume voice applications where latency and cost matter more than peak fidelity.
Both models are new, not a repackaging of earlier work. AI Insiders covered the base MAI-Image-2.5 model in May, when it landed in the Arena leaderboard’s top three for image generation. Pro is a separate, higher-end variant that arrived two months later, and it ships alongside a voice model that has not previously appeared in this newsletter. Microsoft priced Pro at $5 per million text-input tokens, $8 per million image-input tokens, and $106 per million image-output tokens. MAI-Voice-2-Flash costs $15 per million characters. It runs twice as fast as MAI-Voice-2 and is 32 percent cheaper to operate, according to Microsoft.
The naming pattern is worth noting. “Flash” tiers across the industry mark a cost and latency floor rather than a capability ceiling, and that floor is where most production traffic actually runs. A creative team generating a handful of hero images can absorb Pro’s per-token price. A call center handling millions of voice interactions a day cannot, which is why Microsoft built Flash specifically for that job and is routing enterprise voice traffic through it first.
The more consequential story is where these models are already deployed. Bing Image Creator now runs on Microsoft’s own models end to end by default, a shift from configurations that previously drew on outside providers. That switch alone is notable. Microsoft says swapping MAI-Image-2.5 into PowerPoint’s image-to-image feature cut GPU costs by as much as 84 percent compared with GPT-Image-2, OpenAI’s image model. That comparison is the clearest data point yet: Microsoft is substituting its own infrastructure for a partner’s, inside a product that partner also competes to serve directly. OneDrive’s editing tools switched to MAI-Image-2.5 as well, with Microsoft citing a 26 percent increase in save rates and roughly 25 percent lower P95 latency since the change.
Voice tells the same story on the enterprise side. Dynamics 365 Contact Center, the platform behind call center agent deployments at customers including T-Mobile and EasyJet, has switched to MAI-Voice-2-Flash. Microsoft says the change cut GPU costs by up to 89 percent. Azure Voice Live, the developer-facing product for building speech agents, now runs on Flash as well. Both examples point the same direction.
Every figure Microsoft cites here (the 84 percent, the 89 percent, the 26 percent save-rate lift) comes from its own internal comparisons. None of it has been replicated by an outside evaluator, and Microsoft has not published the workload definitions or hardware baselines a third party would need to check the math. The GPT-Image-2 comparison in particular is a company measuring its own product against a competitor’s model, on its own infrastructure, for its own benefit.
None of that undercuts the strategic logic. Every workload Microsoft moves onto MAI is inference spend it no longer pays OpenAI to handle, even as the two companies remain bound together through Azure’s OpenAI hosting agreement. Operators building on Copilot or Azure AI Foundry should treat MAI-Image-2.5-Pro and MAI-Voice-2-Flash as live alternatives to test now, not a distant roadmap item. Watch how many more Microsoft products default to MAI over the next two quarters, since each one marks spend shifting away from a partner Microsoft also competes against.
Announced by Microsoft AI on July 23, 2026.
