Vercel AI Gateway adds five named Microsoft MAI models for voice, transcription and images

The integration saving is in deployment, not in model selection. According to the dev.to write-up of Vercel's announcement, the model call sits in the same application environment as the rest of the project, so no separate deployment process is needed for each modality ^[12]. Five named models ^[14] still means five behaviours to test. Each one has to be tried against the product's own audio, text and images.

Pricing needs more work than a catalog page suggests. The pages publish model-specific prices, but the unit changes with the model type: characters, images or hours of audio ^[6]. That makes the price table closer to three tables than one. A team has to convert its own traffic into each unit before two lines can be compared, for example characters spoken per session, images per user action or audio hours per day. The write-up calls the table useful for early budgeting and not a full project cost estimate ^[8]. The dev.to summary does not include any per-unit price, so the figures sit on Vercel's catalog pages.

The data claim needs the same care. Vercel highlights MAI's zero data retention posture alongside its emphasis on data control and transparency about training usage ^[9]. The write-up adds that this does not remove the need to evaluate what data an application sends to a model ^[10]. In our view, the claim transfers to a product only if it covers the full path from the app through the gateway to the model. That scope is worth confirming in the documentation before real recordings go through.

For a first build, the write-up recommends starting with a single, narrowly scoped job, like transcription of uploaded recordings or image generation for one specific feature. It then says to estimate usage in the published billing unit before exposing the feature broadly ^[13].