How to Install / Upgrade: Fish Audio models now available on Vercel AI Gateway for free

What happened

Fish Audio models now available on Vercel AI Gateway for free

Vercel has added Fish Audio's text-to-speech models to its AI Gateway, giving developers a single unified endpoint through which they can call audio generation alongside the language and image models already available on the platform. The integration means you no longer need a separate Fish Audio account or a custom HTTP client wired to Fish Audio's own API — your existing Gateway credentials and workflow cover it.

To mark the launch, every Fish Audio model on AI Gateway is free for thirty days, through September 18. The standard rate for text-to-speech on the Gateway is fifteen dollars per million characters, so the window is a genuine opportunity to prototype, benchmark, and ship production audio features before the meter starts running.

What shipped

How it works

Fish Audio's audio generation models are now listed in the AI Gateway model catalog. The models handle text-to-speech conversion, taking a string of text and returning synthesized speech audio. Access goes through the same Gateway endpoint you already use for other providers, authenticated with the same API key. No separate Fish Audio credentials are required when you route through the Gateway. The free-tier pricing applies automatically to all requests made before September 18; you do not need to enter a coupon or opt in to a special plan.

What changed for builders

Before this integration, connecting to Fish Audio required maintaining a direct integration: a separate base URL, a separate authentication header, and custom error-handling logic for a provider outside your existing Gateway setup. Now all of that collapses into the Gateway's standard provider-switching mechanism. If your codebase already targets the Gateway, adding Fish Audio audio generation is a matter of changing the model identifier in your request, not restructuring your HTTP client.

The practical consequence is that audio generation becomes composable with your existing Gateway usage patterns. You can log, cache, rate-limit, and observe audio calls through the same dashboard and middleware you already use for text and vision models. Retry logic, spend tracking, and team-level access controls apply uniformly. If you are evaluating multiple text-to-speech providers side by side, the Gateway gives you a single surface for that comparison rather than a patchwork of different clients.

Why it matters

How to install or upgrade

If you are starting fresh, create a Vercel account and navigate to the AI Gateway section of your project dashboard. Enable the Gateway for your project, then generate an API key from the credentials panel. That key is what you will use for all subsequent calls, including Fish Audio models.

If you already use the Gateway, there is nothing to install. Open your project's AI Gateway settings and confirm that the Fish Audio provider appears in the model list. Because providers are surfaced at the platform level, no SDK update or dependency change is required on your end.

To make a text-to-speech request, point your HTTP client at the Gateway's audio endpoint and set the model field to the Fish Audio model identifier shown in the catalog. Pass your text as the input payload according to the request schema documented in the Gateway reference. The Gateway will forward the request, handle authentication with Fish Audio on your behalf, and return the synthesized audio in the response body. During the free window, the usage that would normally accrue toward the fifteen-dollar-per-million-characters rate will be waived automatically.

Who is affected

If you use a framework or SDK that wraps the Gateway, check whether it exposes an audio or speech generation method. The underlying transport is the same, so any SDK that already supports other Gateway providers should be able to call Fish Audio models with minimal adjustment. Verify you are on a version of the SDK that includes the Gateway's audio endpoint support before writing integration code.

For teams managing multiple environments, set the Gateway base URL and API key as environment variables and scope them per environment in Vercel's project settings. This keeps credentials out of source control and lets you test against the free tier in development without touching production configuration.

Gotchas and compatibility

The free period ends on September 18. Requests submitted on or after that date will be billed at the standard rate, so budget accordingly if you plan to keep audio features in production after the window closes. There is no grace period or partial-month proration mentioned, so treat September 18 as a hard cutoff.

What to watch next

Audio generation has different latency and payload size characteristics than text completion. If you have aggressive timeout values set in your Gateway configuration or your HTTP client, test them with audio responses before going to production. Audio payloads are larger than text payloads, and synthesis time scales with input length, so a timeout that is fine for a chat completion may be too short for a longer speech output.

The integration routes through the Gateway, which means any Gateway-level quotas or rate limits your plan includes apply to audio calls too. If you are on a plan with a request-per-minute ceiling, high-volume audio generation will count against that ceiling alongside your other model calls.

What to watch next

Watch the Vercel changelog and the AI Gateway model catalog for updates to the Fish Audio model lineup and for any changes to the post-September-18 pricing structure. The free period is also a practical opportunity to measure real-world latency and audio quality against your product requirements, so use this time to run evaluations with representative input text rather than waiting until the billing window opens.