Vercel Makes Fish Audio models on AI Gateway Free for 30 Days.

Vercel is adding Fish Audio’s voice generation and transcription models to AI Gateway, giving developers another route to build audio features without integrating directly with a separate model provider. All four models are free through September 18, 2026, as part of a 30-day launch promotion.

The changelog announcement covers both text-to-speech and speech-to-text workloads. That makes the release relevant to builders working on narrated content, voice interfaces, multilingual applications, transcription tools and applications that need timestamped audio analysis.

The temporary pricing is straightforward:

• Text-to-speech: Usually $15 per million characters, free through September 18
• Speech-to-text: Usually $0.36 per hour of audio, free through September 18

The promotion removes model usage charges during the stated period, but Vercel has not described it as a permanent free tier. Developers evaluating the integration should therefore account for the regular rates before placing it into a production workflow.

Vercel says four Fish Audio models are available through AI Gateway. Its announcement details three of them, including Fish Audio’s latest text-to-speech model:

• fish-audio/s2.1-pro: A text-to-speech model designed for low-latency streaming. It can clone a voice from a reference recording.
• fish-audio/transcribe-1: A transcription model that returns text, the duration of the source audio and timestamped segments. Its timestamps can extend down to individual words.
• fish-audio/s2-pro: A text-to-speech model covering roughly 80 languages.

The low-latency focus of s2.1-pro could be useful for interactive applications where generated speech needs to begin before a complete response has been produced. Voice cloning also opens up more customized output, although builders will need appropriate permission for any reference voice they use.

transcribe-1 offers more than a plain transcript. Duration data and word-level timestamps can support searchable recordings, synchronized captions, meeting summaries and interfaces that let users jump to a specific passage in an audio file.

AI Gateway is Vercel’s model access layer. Adding Fish Audio means developers can bring audio models into the same Vercel-centered stack they may already use for deploying and operating AI applications, rather than treating voice as an entirely separate infrastructure decision.

For small teams, the immediate benefit is room to test. A builder could prototype spoken responses, generate multilingual narration or assess transcription accuracy during the free period before committing to the normal per-character or per-hour economics.

The update also broadens Vercel’s AI tooling beyond text-based model interactions. It follows other efforts to make application capabilities available through its platform, including the company’s recent Instagram adapter for Chat SDK Direct Messages and its move to bring production-aware deployment actions into Slack.

There are still practical questions developers will need to answer in testing. Streaming latency, language quality and transcription accuracy will depend on the workload and source audio. The promotional window gives teams until September 18 to measure those factors without paying Fish Audio model usage fees through AI Gateway.

What is Vercel Fish Audio models on AI Gateway?

It is Vercel’s integration of four Fish Audio voice and audio models into AI Gateway. Developers can use the available models for text-to-speech, voice cloning and speech transcription through Vercel’s model access layer.

How long are the Fish Audio models free on Vercel AI Gateway?

Vercel says every available Fish Audio model is free on AI Gateway for 30 days, through September 18, 2026. After the promotion, the announced regular rates are $15 per million text-to-speech characters and $0.36 per hour of speech-to-text audio.

Which Fish Audio model supports word-level timestamps?

fish-audio/transcribe-1 returns a transcript alongside audio duration and timestamped segments. Vercel says the timestamps can be provided down to individual words, which can help with captions and navigable transcripts.