What happened
On October 27, 2026, Vercel announced a partnership with ElevenLabs that brings 14 ElevenLabs models to AI Gateway. Vercel says every ElevenLabs model is available at 50% off through November 9. The gateway groups the new catalogue into six audio categories: speech to text, streaming text to speech, realtime voice agents, music, sound effects, and multi-voice dialogue.
The speech-to-text options are elevenlabs/scribe-v2 and elevenlabs/scribe-v2-realtime. Streaming voice output is available through elevenlabs/eleven-flash-v2.5, elevenlabs/eleven-flash-v2, elevenlabs/eleven-multilingual-v2, and elevenlabs/eleven-v3-conversational. For a single realtime conversational model, the gateway offers elevenlabs/eleven-realtime-v1-max and elevenlabs/eleven-realtime-v1-mini. The remaining catalogue is elevenlabs/music-v2.5, elevenlabs/music-v2, elevenlabs/sound-effects-v2, elevenlabs/eleven-v4, elevenlabs/eleven-v4-turbo, and elevenlabs/eleven-v3.
Vercel describes two different ways to construct a voice interaction. In a composable pipeline, Scribe v2 Realtime transcribes a 16 kHz microphone stream while a user is speaking. A separately chosen language model writes the response, and Eleven Flash v2.5 streams the response aloud before the text model has finished generating. The company’s example uses streamTranscribe, streamText, and experimental_streamSpeech. Its stated advantage is modularity: transcription, language reasoning, and voice generation can be swapped without changing the rest of the pipeline.
For applications that want one live spoken session instead, the two Eleven Realtime models receive user audio and return spoken replies over one WebSocket connection. Vercel says the server mints a short-lived token, allowing the browser to establish the session without exposing an API key. Its useRealtime hook in @ai-sdk/react handles the WebSocket connection, microphone capture, and audio playback. The models support semantic voice activity detection, which determines when a user has finished speaking, but do not support server-side VAD.
The announcement also extends the gateway beyond conversational speech. Music v2.5 and Music v2 generate tracks from textual instructions, with a requested duration and optional vocals; Vercel’s example asks for a 30-second lo-fi track through generateSpeech with musicLengthMs: 30000. Sound Effects v2 produces shorter effects and looping ambience. Eleven v4, Eleven v4 Turbo, and Eleven v3 can render a script containing several assigned voices as a continuous recording rather than separate clips.
All requests use one gateway API key and appear in AI Gateway’s usage, cost, and trace views alongside other traffic. Vercel says AI Gateway provides a unified model API as well as retries, failover, performance optimizations, custom reporting, Zero Data Retention support, API-key budgets, and routing rules. It says provider pricing is reflected without inference markup or a platform fee, including for bring-your-own-key requests.
What the archive adds
This is a new capability on a known gateway story. Earlier notes covered Microsoft AI audio models arriving in AI Gateway and Browserbase Search and Fetch becoming gateway tools. The ElevenLabs release broadens the same gateway surface into a more complete set of audio production and interaction primitives.
Trend signal
The announcement supports Agent-ready software exposes capabilities. Rather than requiring teams to negotiate separate provider integrations for each audio function, AI Gateway exposes named, selectable capabilities—transcription, speech, realtime sessions, and generation—through a common operational boundary. The source establishes the interface and observability layer, not whether those capabilities are equally reliable or safe in every deployment.
Why now
The timing is explained only as a launch promotion: Vercel is offering all ElevenLabs models at half price through November 9. The practical timing for builders is therefore short: teams evaluating an audio workflow can test usage patterns and integration mechanics while the temporary discount applies.
Connections to other events
AI Gateway had already added Browserbase Search and Fetch as tools available across supported model providers. This release adds audio inputs and outputs to the same broader gateway approach. The prior Microsoft AI integration also concerned managed audio access; ElevenLabs adds realtime voice, music, effects, and multi-voice generation rather than only speech generation and transcription.
Why it matters
Voice product work is often assembled from multiple provider-specific APIs: a speech recognizer, a text model, a speech synthesizer, session plumbing, client-side audio handling, and separate monitoring. Vercel’s announcement does not remove the application design work, but it offers a common access and observation layer around more of that stack. A team can choose the composed route when it needs to replace individual stages, or the realtime route when a single WebSocket voice session is a better fit.
That choice matters operationally as much as it does at the demo layer. Audio traffic can now sit in the same usage, cost, and trace views as language-model traffic, according to Vercel. Teams evaluating voice agents should treat that as an opportunity to measure the full interaction path—rather than only comparing a speech model’s isolated output quality.
What happens next
- Teams will test composed speech pipelines against single-model realtime sessions for different latency and control requirements. (3 months, 70%). Watch for: product documentation or examples that compare these architectures.
- AI Gateway will receive additional provider capabilities beyond text generation and web retrieval. (6 months, 60%). Watch for: new model categories or tool integrations announced by Vercel.
Limits and uncertainty
The source is a Vercel launch announcement and describes supported model IDs, SDK patterns, gateway features, and a promotional price. It does not provide comparative latency, quality, reliability, regional availability, concurrency limits, or per-model pricing. It also does not establish that the suggested pipelines will meet the safety, privacy, consent, accessibility, or telephony requirements of a particular deployment. “Higher-than-provider uptime” is presented as a gateway capability, not accompanied here by a measured availability result. Builders should validate end-to-end behavior, costs, and data-handling terms in their own intended configuration.
Skill to build now
Build a small voice-path benchmark rather than a general demo. Send the same short microphone input through the modular route—Scribe v2 Realtime, a selected text model, and Eleven Flash v2.5—and record transcription completion, first audible response, final response, cost, and trace data. Then repeat the task with an Eleven Realtime model. This makes the architectural choice concrete: choose modular stages when substitution and inspection matter, and assess the single-session route when conversational flow is the priority.
Reader takeaway
Vercel’s AI Gateway now offers a much wider ElevenLabs catalogue through its common API and operating surface. The important decision is not simply which voice sounds best: builders can now compare a swappable transcription-to-language-to-speech pipeline with a single realtime voice session while monitoring both through the same gateway.
Open question
Will the gateway’s common usage and trace layer give teams enough end-to-end visibility to manage audio latency and cost across these different architectures?