What shipped? Fish Audio's speech and transcription models are now available through Vercel AI Gateway, free for 30 days.
What changed? Four models landed: fish-audio/s2.1-pro (voice cloning TTS, built for low-latency streaming), fish-audio/s2-pro (80+ languages with inline speech tags), fish-audio/s1 (emotion and sound-effect markers), and fish-audio/transcribe-1 (word-level timestamps). During the free period (through September 18), all usage costs zero. After that, TTS is $15/million characters and STT is $0.36/hour of audio. Append -free to the model name (e.g., fish-audio/s2.1-pro-free) to auto-stop when the offer ends. Available via AI SDK v7's generateSpeech() and transcribe(), plus the Gateway playground in your browser.
Why does a builder care? Speech is the last modality Vercel's AI Gateway was missing — now every major input/output path goes through one API key, one bill, one dashboard. The 30-day free trial costs nothing to evaluate, and the -free suffix is a nice protection against accidental post-period billing. For agent builders: voice I/O through the same gateway as your LLM calls means unified observability and no separate provider accounts to manage.