Vercel Adds Gemini 3.5 Transcribe to AI Gateway.

Vercel is expanding AI Gateway beyond text and image generation with support for Gemini 3.5 Transcribe, Google’s new speech-to-text model. Developers can now process completed recordings or stream microphone audio through the same gateway used to access other AI models.

The integration, announced in Vercel’s changelog, gives application builders two distinct transcription options. One handles an entire audio file in a single request, while the other returns an evolving transcript over a WebSocket as someone is speaking.

That distinction makes the update relevant to a broad range of products. Batch transcription can support meeting archives, uploaded interviews, podcasts and customer-call analysis. Live transcription is better suited to captions, voice interfaces and applications that need to respond before a recording has ended.

Google’s Gemini 3.5 Transcribe is available through AI Gateway under two model identifiers:

• google/gemini-3.5-transcribe accepts a complete recording and returns its transcription in one request.
• google/gemini-3.5-transcribe-live receives audio over a WebSocket and updates the transcript while the recording continues.

The live implementation uses streamTranscribe, a capability introduced in AI SDK V7. It opens the WebSocket and accepts a ReadableStream containing raw audio chunks, allowing developers to feed microphone input directly into the transcription workflow. Vercel says developers should install the latest versions of both the ai package and @ai-sdk/gateway to use streaming transcription.

The model automatically detects the language being spoken and supports more than 85 languages. It can also continue following a speaker who changes languages during a recording, a useful capability for multilingual meetings, interviews and support conversations.

Custom vocabulary provides another practical control. Developers can supply names, specialist terminology, product jargon and preferred spellings that might otherwise be misunderstood by a general-purpose transcription model. For small businesses, that could reduce the cleanup required when processing recordings containing brand names, industry acronyms or customer-specific terminology.

The model itself is only part of Vercel’s pitch. By routing Gemini 3.5 Transcribe through AI Gateway, developers can call it with one gateway API key rather than maintaining a separate provider account and integration. Vercel also lists automatic fallbacks, spend tracking and request traces among the gateway-level capabilities available to users.

Centralising those controls may be particularly useful for teams already using AI Gateway for other model types. It creates a common operational layer for monitoring requests and expenditure, while reducing the amount of provider-specific infrastructure an application needs to maintain.

The release also shows Vercel broadening the gateway into a more general AI application layer. On the same date, the company added Meta’s image model, as covered in Vercel Adds Muse Image Generation and Editing to AI Gateway. Together, the releases give builders access to text, audio and image workflows behind a more consistent integration surface.

There are still implementation choices to make. Builders need to decide whether users can wait for a completed transcript or require immediate partial results. Live transcription also introduces a persistent WebSocket and a stream of audio chunks, making it technically different from uploading a finished file through a conventional request.

Vercel did not announce separate pricing for Gemini 3.5 Transcribe in the update. Its practical value will therefore depend on the wider AI Gateway pricing and model usage terms, as well as the latency and accuracy developers observe with their own audio, languages and specialist vocabulary.

What is Vercel Gemini 3.5 Transcribe?

Gemini 3.5 Transcribe is Google’s audio-to-text model, now accessible through Vercel AI Gateway. It can transcribe either a complete recording or a live audio stream, with automatic language detection and support for custom vocabulary.

What is the difference between the two transcription variants?

google/gemini-3.5-transcribe processes a complete recording in one request. google/gemini-3.5-transcribe-live uses a WebSocket to return transcript updates while audio is still being recorded.

When is Gemini 3.5 Transcribe available on Vercel AI Gateway?

Vercel announced its availability on August 26, 2026. Streaming transcription requires AI SDK V7 and the latest AI SDK Gateway package.