OpenAI ships prompt-to-voice: spin up a GPT-Live voice from text
No reference clip, no voice actor: describe it, get an ID, build with it. Plus Cursor's Environments API, Fish Audio's open TTS, and a 2B-active local model.
Copy markdown
Describe the voice, don't record it
POST to /v1/audio/voices with type=prompt and a plain-text description, and you get back a reusable voice ID. Creating one is free (100 per org) and needs no reference clip or consent recording. It works only in GPT-Live, billed at normal session rates, so a branded voice agent no longer needs a studio booking.
Nine new GPT-Live voices ship alongside
OpenAI also dropped nine natural voices for GPT-Live, including Nuri, Juni, Flitz (German) and Noeul (Korean), plus accent improvements across French, German, Hindi, Mexican Spanish, Korean and Japanese. More languages to prototype voice agents in, today.
Cursor opens its Environments API
Cursor's cloud agents can now build interactive canvases you share by link, and the Environments API enters beta: up to 300 repositories per environment, loading 5x faster. Enough to wire a whole monorepo set into background agents at once.
Fish Audio Drama 3: direct open TTS like a script
A 35B open-weight text-to-speech model you steer with script-style stage directions: tone, pacing and emotion per line. The weights are on Hugging Face, so the voice layer can run on your own box instead of a hosted endpoint, the open counterpart to OpenAI's launch.
Liquid LFM2.5 runs agents at 2B active
A 24B mixture-of-experts that fires only ~2B parameters per token, with 131K context and ~184 tok/s on an M5 Max, shipping in GGUF, MLX and vLLM under a free-weights license. Fast tool-calling loops that stay on the laptop; the hosted API runs $0.10 / $1.00 per million in/out.