For AI agents: a documentation index is available at the root level at /llms.txt. Append /llms.txt to any URL for a page-level index, or .md for the markdown version of any page.
GPT-Live voices: Nine new GPT-Live voices are available: Nuri, Juni, and Flitz (German), Brise and Sillage (French), Nira (Hindi), Harema and Shida (Japanese), and Noeul (Korean). Assistants that don’t use a GPT-Live model can no longer be saved with a GPT-Live only voice.
Features
GPT-Live prompt editor: The GPT-Live speaker and reasoner prompts now use the rich text editor, with visual and code views and a fullscreen mode.
Reasoning effort for structured outputs: You can now set a reasoning effort for structured outputs that use an OpenAI reasoning model.
Deepgram feedback filter setting: Set feedbackFilterEnabled to false on a Deepgram transcriber to keep caller speech that repeats what the assistant just said. It’s on by default.
Simulation suite generation API: You can now draft requirements and generate simulation suites through the API with a private API key.
Fixes and Improvements
ElevenLabs pronunciation dictionaries: Calls that use an ElevenLabs pronunciation dictionary without a pinned version no longer fail. They use the dictionary’s latest version.
ElevenLabs silent turns: When ElevenLabs returns no audio for a turn, the call retries once and then switches to your fallback voice.
GPT-Live phone audio: GPT-Live phone calls no longer have short gaps of silence in the middle of words.
