Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableVercel Blog→ original

Vercel adds 14 ElevenLabs models to AI Gateway with 50% discount for 14 days

Vercel has added 14 ElevenLabs models to AI Gateway, so a voice app can now be built with a single API key. A user speaks, a speech model transcribes it, any language model writes the reply, and an ElevenLabs voice reads it out as it is generated. Until November 9, all of these models in the gateway cost 50% less, which makes it a convenient moment to test voice in a product.

AI-processed from Vercel Blog; edited by Hamidun News
Vercel adds 14 ElevenLabs models to AI Gateway with 50% discount for 14 days
Source: Vercel Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceVercel Blog ↗
Editor-in-Chief
Six typesspeech recognition, text-to-speech, voice agents, music, sound effects, dialogue
Key metric #2Eleven Realtime models keep a conversation in a single WebSocket session without separate steps
Key metric #3Vercel adds no markup and no platform fee, including for requests that use your own key

A voice interface usually ends up costing more than it seems at the start. You need one service for speech recognition, a second for the language model, a third for speech synthesis, and each has its own key, its own bills and its own logs. Vercel wants to clear up that sprawl: AI Gateway now offers 14 ElevenLabs models, all accessible through one API key. To mark the launch, a 50% discount applies to all ElevenLabs models in the gateway through November 9 inclusive.

What was actually connected

There are six types of models. Speech recognition (Scribe v2 and Scribe v2 Realtime) turns voice into text, including while the person is still speaking. Streaming speech synthesis (Eleven Flash v2.5, Flash v2, Multilingual v2, v3 Conversational) starts playing audio while the language model is still finishing its answer. Voice agents (Eleven Realtime v1 Max and v1 Mini) handle the entire conversation within a single model. Separately, there is music generation (Music v2.5 and v2), sound effects (Sound Effects v2), and multi-voice dialogue synthesis (Eleven v4, v4 Turbo, v3).

The practical benefit is that all requests land in the same spending, usage and tracing reports as the rest of the traffic through AI Gateway. You don't need a separate account for audio or to reconcile invoices by hand.

Two ways to build a voice assistant

The first approach is a three-step pipeline. Take the example from the documentation: a passenger asks whether they will make a connection. Scribe v2 Realtime transcribes the speech on the fly, then the language model you choose (in the example, an Anthropic model from the Claude Haiku line) combines the flight status and the distance to the boarding gate, and Eleven Flash v2.5 reads the answer as it appears. The user hears the start of the sentence without waiting for the full text. Each step is a separate model, so recognition, the "brain" or the voice can be replaced without touching the rest of the code.

The second approach is one model for the whole dialogue. Eleven Realtime v1 Max and v1 Mini take the user's audio and respond with voice over a single WebSocket session. You don't need to connect recognition and synthesis separately. This suits a support line or an in-app assistant where natural interruptions and a live pace matter: to the question "Can I change the delivery address?", the agent immediately replies "Sure, tell me your order number."

The security scheme is simple: the server only issues a short-lived token for the chosen model, and the browser opens the connection itself. That way the API key never reaches the client. On the frontend, the useRealtime hook from @ai-sdk/react handles the microphone, the connection and playback. One important detail: Realtime models detect the end of an utterance only through semantic pause detection (semantic-vad); the server-vad mode is not supported.

Music, effects and dialogue

Music v2.5 and v2 build a finished track of the required length from a text description, with or without vocals. You can specify genre, mood and instruments, and for more control describe a song section by section, each with its own style and duration. Sound Effects v2 does the same for short sounds: footsteps, a door slam, rain looped for a game level. In code it is the same generateSpeech call, except the input is a description of the sound rather than text to be read, for example "calm lo-fi beat with soft piano" for 30 seconds.

The Eleven v4, v4 Turbo and v3 models voice a script with several characters in a single request. Each line is assigned a voice, a voice change starts a new turn, and the output is one continuous recording. For a two-host podcast or a full-cast scene, you don't have to stitch together dozens of separate clips.

Pricing and what to watch for

According to Vercel, AI Gateway shows provider prices without markup and charges no platform fee for inference, including when you use your own key (BYOK). It also offers retries, fallback to a backup option, budgets per key, routing rules and support for Zero Data Retention.

There are limitations too. Some SDK functions are marked experimental (for example, streamSpeech and streamTranscribe), so the interface may change. The discount is temporary, and Vercel does not publish data on latency or quality compared with connecting directly to ElevenLabs. Teams already paying ElevenLabs directly should compare the final invoices.

The takeaway for practitioners: if you are building a voice assistant, a support bot or a game with sound, these two weeks are an inexpensive way to test the idea. A prototype can be assembled in an evening within a single SDK, and the decision on a permanent setup can wait until you have compared latency and cost on real conversations.

Story

Key Questions Answered

What was actually connected?

There are six types of models. Speech recognition (Scribe v2 and Scribe v2 Realtime) turns voice into text, including while the person is still speaking. Streaming speech synthesis (Eleven Flash v2.5, Flash v2, Multilingual v2, v3 Conversational) starts playing audio while the language model is still finishing its answer. Voice agents (Eleven Realtime v1 Max and v1 Mini) handle the entire conversation within a single model.

What are the key details on Two ways to build a voice assistant?

The first approach is a three-step pipeline. Take the example from the documentation: a passenger asks whether they will make a connection. Scribe v2 Realtime transcribes the speech on the fly, then the language model you choose (in the example, an Anthropic model from the Claude Haiku line) combines the flight status and the distance to the boarding gate, and Eleven Flash v2.5 reads the answer as it appears. The user hears the start of the sentence without waiting for the full text.

What are the key details on Music, effects and dialogue?

Music v2.5 and v2 build a finished track of the required length from a text description, with or without vocals. You can specify genre, mood and instruments, and for more control describe a song section by section, each with its own style and duration. Sound Effects v2 does the same for short sounds: footsteps, a door slam, rain looped for a game level.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…