Vercel brings 14 ElevenLabs audio models to AI Gateway with 50% launch discount

Vercel and ElevenLabs have partnered to bring 14 ElevenLabs audio models to AI Gateway, Vercel's unified API for calling AI models. To mark the launch, every ElevenLabs model on the gateway is half price for the first 14 days, with the discount running through November 9.

What launched

According to Vercel, the gateway now covers six categories of ElevenLabs audio models, all reachable with a single API key:

• Speech to text, for transcribing recordings or captioning live audio as it happens
• Streaming text to speech, which starts playing audio while a language model is still composing its reply, so users hear the beginning of an answer before generation finishes
• Realtime voice agents that hold a live spoken conversation with a user in one session — Vercel points to support lines, tutors and in-app assistants as natural fits
• Music generation, where a plain-text description returns a finished song or instrumental at a requested length
• Sound effects, covering game and interface sounds as well as ambience that can loop
• Multi-voice dialogue, where a conversation voiced by several different speakers comes back in a single call

Vercel says requests through the gateway appear in the same usage, cost and trace dashboards as a developer's other AI Gateway traffic.

Composable speech pipelines

Much of the announcement details how the pieces fit together. Scribe v2 Realtime converts a user's speech into text while they are still talking; a language model of the developer's choosing then drafts an answer, and Eleven Flash v2.5, a low-latency voice model, reads that answer aloud as it streams in. Because transcription, reasoning and speech are separate steps, each one can be swapped out without disturbing the rest of the pipeline.

Realtime voice agents

For conversations where natural back-and-forth matters, Eleven Realtime v1 Max and Eleven Realtime v1 Mini handle the entire spoken exchange inside one model over a single WebSocket session, with no separate transcription or speech steps to wire up. Vercel describes a browser setup in which the server's only job is to mint a short-lived token for the chosen model, keeping the API key off the client. These models rely on semantic voice activity detection to judge when a user has finished speaking and do not support server-side VAD.

Music, sound effects and multi-voice dialogue

Music v2.5 and Music v2 generate original tracks from a text description — genre, mood, instruments, length, with or without vocals — and songs can be laid out section by section, each with its own style and duration. Sound Effects v2 targets shorter audio such as footsteps, doors slamming or rain that can loop seamlessly beneath a game level.

On the dialogue side, Eleven v4, Eleven v4 Turbo and Eleven v3 perform a multi-speaker script in one request. Developers assign a voice to each line, and a two-host podcast or a full-cast scene arrives as one continuous recording with nothing to stitch together. Script parts can also stream in from a language model that is writing the dialogue on the fly.

Pricing and platform details

Vercel positions AI Gateway as a single API for model calls, with usage and cost tracking, retries, failover and performance optimizations that the company says deliver higher uptime than connecting to providers directly. The service also offers custom reporting, Zero Data Retention support, per-key budgets and routing rules. Pricing matches what the provider charges, with no added markup and no platform fee on inference, including on Bring Your Own Key requests.

Why it matters

Audio has often been the awkward piece of the AI stack: developers stitching together transcription, reasoning and speech typically juggle several provider keys and billing accounts. Placing 14 ElevenLabs models behind one gateway key — with the same observability, budgeting and failover tooling already used for text models — makes voice a first-class part of the pipeline rather than a bolt-on. The 50% introductory pricing through November 9 lowers the barrier to experimenting. For teams already building agents on Vercel's platform, a spoken assistant, a generated soundtrack or a multi-voice podcast can now be assembled from the same API surface as the language models driving them.