Mutable alias for xAI's newest speech-to-speech voice agent model; since October 14, 2026 it points to Grok Voice 3.0, after earlier tracking Voice 1.0 and 2.0. Runs over a realtime, OpenAI Realtime-compatible WebSocket in 20+ languages.
July 2026 update to gpt-realtime-2 with better alphanumeric recognition, silence and noise handling, and interruption behavior; configurable reasoning effort, 128K context. OpenAI's named replacement for gpt-realtime and gpt-4o-realtime.
Distilled, lower-cost version of gpt-realtime-2.1 released July 2026, with configurable reasoning and a 128K context at $10/$20 per 1M audio tokens in/out. OpenAI's named replacement for gpt-realtime-mini, tts-1 and gpt-4o-mini-tts.
Gemini 3.5 Live Translate is a June 2026 Live API preview for continuous speech-to-speech translation across 70+ languages, detecting the source language and keeping the speaker's intonation, pacing and pitch. It remains Google's dedicated live-translation model, with no shutdown announced.
Realtime API speech-to-speech model released May 2026, adding configurable reasoning effort and tool use for voice agents, with a 128K context and 32K max output tokens. Updated by gpt-realtime-2.1 in July 2026.
Grok Voice 1.0, xAI's April 2026 speech-to-speech agent model with background reasoning, tool calling and 25+ languages. Superseded by Grok Voice 2.0 in July and Grok Voice 3.0 in October 2026; no longer in xAI's voice model table.
Gemini 3.1 Flash Live, a March 2026 Live API preview for real-time audio-to-audio conversation with thinking, Search grounding and function calling. Legacy since Gemini 3.8 Live reached GA in September 2026; shutdown is set for November 17, 2026.
OpenAI's February 2026 speech-to-speech model for Chat Completions, taking text or audio and replying with audio and text, with a 128K context. Successor to gpt-audio and gpt-4o-audio-preview; works per request rather than as a live session.
February 2026 update to gpt-realtime for low-latency voice agents, with OpenAI-reported gains in instruction following, tool calling and alphanumeric transcription; 32K context. Still offered, but gpt-realtime-2 and 2.1 since added reasoning and a 128K context.
December 2025 Live API snapshot of Gemini 2.5 Flash native audio for real-time voice dialogue, with improved function calling and instruction following. Superseded by Gemini 3.1 Flash Live and now Gemini 3.8 Live; shutdown is scheduled for November 17, 2026.
Qwen3-Omni-Flash as a speech-to-speech model: takes text, image, audio or video and replies with text and generated speech, covering 119 text and 20 speech languages (December 2025 build). Superseded by Qwen3.5-Omni (March 2026).
Amazon's December 2025 speech-to-speech model on Bedrock for live voice conversations over a bidirectional stream, with barge-in, adjustable pause sensitivity and async tool calls in seven languages; 1M-token context. Superseded by Nova 2.5 Sonic on 2026-10-05.
OpenAI's first generally available Realtime API speech-to-speech model (snapshot gpt-realtime-2025-08-28), which launched with image input and the Marin and Cedar voices. Deprecated in July 2026; shuts down 2027-01-20, replaced by gpt-realtime-2.1.
A provider bundle: Deepgram's Voice Agent API set to Nova-3 for listening, OpenAI GPT-4o mini for reasoning and the Aura-2 Thalia English voice. One WebSocket handles turn-taking, barge-in and function calls; the API reached general availability in June 2025.
Hume's third-generation Empathic Voice Interface, one speech-language model that transcribes, responds and speaks with prompt-designed or cloned voices. It remains the default EVI version alongside EVI 4-mini, but Hume closes the EVI API on November 13, 2026.
Inworld's Realtime API, an OpenAI Realtime-compatible speech-to-speech service, set up here with Inworld STT-1 for transcription, GPT-4o for replies and Inworld TTS-2 for voice, with turn detection, interruptions and tool calling over one WebSocket session.
