Google's Gemini 3.5 Transcribe models arrive on Vercel's AI Gateway

Google transcription models land on Vercel's gateway

Vercel has added Google's Gemini 3.5 Transcribe speech-to-text models to AI Gateway, its unified routing layer for model calls. According to a changelog post on the Vercel blog dated August 26, developers get two ways to turn audio into text: a batch model for finished recordings and a live model that streams a transcript over a WebSocket while the audio is still coming in.

Two variants, two workflows

The first variant, google/gemini-3.5-transcribe, handles a complete recording in a single request. The second, google/gemini-3.5-transcribe-live, keeps a socket open and continuously updates the transcript as the recording continues — the pattern suited to live captioning, meeting notes or voice-driven interfaces.

Vercel says the model detects the spoken language on its own, covers more than 85 languages, and can follow a speaker who switches languages partway through. Developers can also supply custom vocabulary so that names, domain jargon and unusual spellings are rendered correctly.

Streaming transcription lands in AI SDK V7

According to the post, streaming transcription is new in AI SDK V7 and becomes available after installing the latest ai and @ai-sdk/gateway packages. The streamTranscribe function opens the socket and accepts a ReadableStream of raw audio chunks, which means a microphone stream can be piped straight through. The expected audio format is declared with inputAudioFormat; the documented example uses 16 kHz 16-bit PCM. A Google-specific provider option selects between two transcription modes, SMART and VERBATIM, with VERBATIM as the default.

For audio already stored on disk, the transcribe function sends the file in one request and returns the resulting text. In the sample code, an MP3 is read from the filesystem and its transcript is printed from the response.

Both functions are currently exported under an experimental_ prefix, which signals that the API surface may still change as the SDK evolves. For anyone who wants to evaluate the model before writing code, Vercel also points to a browser-based playground, Gemini 3.5 Transcribe Live, where audio can be sent directly and the transcript read in the page.

Pricing and platform context

Vercel positions AI Gateway as a single API for calling models, tracking usage and cost, failing over between providers, and applying performance optimizations that the company claims deliver higher uptime than the underlying providers. The gateway also ships with custom reporting, per-key budgets and routing rules.

On cost, Vercel states that the gateway mirrors provider pricing with no markup and charges no platform fee on inference, including on Bring Your Own Key requests — meaning developers route through Vercel without paying a premium over Google's own rates.

Why it matters

Real-time transcription is a core building block for captioning tools, meeting assistants and voice agents, and until now it often required wiring up a separate speech provider alongside whatever LLM API an application already used. Folding Google's transcription models into the same gateway as other AI calls means one integration, one bill and one place to set failover and budgets.

The live WebSocket variant is the more strategically interesting half of the release: an updating transcript while audio is still being captured is what interactive, low-latency voice products need, and the combination of automatic language detection across 85-plus languages with mid-utterance language switching targets multilingual audiences that many transcription services handle poorly.

There are caveats worth noting. The details come from a single vendor announcement, so independent benchmarks of accuracy, latency and real-world language-switching behavior are not yet available, and the SDK functions are explicitly experimental. Still, for teams already building on Vercel's AI stack, the absence of an inference markup removes the main pricing argument against routing transcription through the gateway rather than calling Google directly.