Fish Audio's cheap voice models land on Vercel's AI Gateway
One key, no separate account, ~11x-cheaper TTS than ElevenLabs — enough to ship a realtime voice agent this weekend. Plus DFlash 2 speeds local inference.
Copy markdown
One key, no Fish account, free 30 days
Fish Audio's OpenAudio S1 and S2 voice models are now callable straight through Vercel's AI Gateway — no separate provider signup, free for the first 30 days, with usage and spend tracked in the dashboard you already use.
~11x cheaper than ElevenLabs
S2 Pro runs about $15 per 1M characters against roughly $120 on ElevenLabs' Multilingual v2 — a blog-length clip is about $0.13. For voice-heavy apps that turns TTS from a real line-item into a rounding error.
Ship a realtime voice agent this weekend
The Gateway already handles realtime voice, TTS and transcription via AI SDK 7's useRealtime hook, with your key staying server-side. Point it at Fish Audio and you have a live, low-latency talking agent with no separate audio vendor to wire up.
Elsewhere: DFlash 2 speeds local inference
Inco open-weighted new speculative-decoding drafters for Qwen3.8-27B and Meta's Muse Glimmer-30B: 2.7-4.6x the throughput of plain autoregressive decoding, drop-in for vLLM, SGLang, llama.cpp and MLX, with identical outputs and no retraining.