Skip to main content
Model details on this page come from each provider’s official documentation and announcements, checked on October 16, 2026. Providers update models, benchmarks, language support and availability often. Always check the linked official source for the latest information before choosing a model.
Realtime (speech-to-speech) models take the caller’s audio in and produce spoken audio out with a single model. They replace the separate speech-to-text, LLM and text-to-speech steps of a pipeline voice agent.

​
At a glance

​
OpenAI

​
Specifications

Values are in tokens and come from each model’s page on developers.openai.com.

​
GPT Realtime 2.1 Mini

gpt-realtime-2.1-mini · Official docs “A distilled reasoning model for faster, lower-cost realtime voice interactions.” OpenAI says it improves alphanumeric recognition over GPT Realtime 2. It supports function calling.

​
GPT Realtime 2.1

gpt-realtime-2.1 · Official docs “GPT-Realtime-2.1 updates GPT-Realtime-2 with improved alphanumeric recognition, silence and noise handling, and interruption behavior.” It supports speech-to-speech with configurable reasoning effort, instruction following and tool use. OpenAI notes that higher reasoning effort can increase latency.

​
GPT Realtime 2

gpt-realtime-2 · Official docs A speech-to-speech model with configurable reasoning effort, “stronger instruction following, and more reliable tool use”. OpenAI’s announcement describes it as bringing “GPT-5-class reasoning” to realtime voice. GPT Realtime 2.1 is an update to this model.
OpenAI publishes GPT Realtime 2’s benchmarks (Big Bench Audio and Audio MultiChallenge) only as charts in its announcement, and publishes no figures for GPT Realtime 2.1 Mini. The Third-party figures above are published by Artificial Analysis, an independent benchmarking company, not by OpenAI. Accessed October 16, 2026.

​
GPT Live 1

gpt-live-1 · Official docs “GPT-Live 1 is a full-duplex voice model for real-time conversations.” It can listen and speak at the same time. It handles pauses, interruptions and backchannels, and hands off reasoning and tool use to a backend agent. It comes with 12 voices.
OpenAI’s Tau3 result pairs GPT Live 1 with GPT-6 Astra (medium reasoning effort) as the backend model. Results with a different backend model will differ.

​
Google

​
Specifications

Values come from each model’s page on ai.google.dev and from Gemini deprecations. Google lists gemini-3.8-live as the replacement for both shutting-down models.

​
Gemini Live 2.5 Flash

gemini-2.5-flash-native-audio-preview-12-2025 · Official docs Provides “low-latency, real-time voice and video interactions with Gemini 2.5 Flash” through the Live API. It supports function calling, search grounding and thinking. Knowledge cutoff: January 2025.
Google lists a shutdown date of November 17, 2026 for this preview model on the Gemini API, with Gemini Live 3.8 as the replacement.

​
Gemini Live 3.1 Flash

gemini-3.1-flash-live-preview · Official docs “A legacy preview model for real-time dialogue and voice-first AI applications.” It supports thinking levels (minimal by default, up to high) and synchronous function calling.
Google lists a shutdown date of November 17, 2026 for this preview model on the Gemini API, and recommends updating to Gemini Live 3.8.

​
Gemini Live 3.8

gemini-3.8-live · Official docs Google calls it the “default option for most low-latency voice agent experiences”. It adds interleaved reasoning and asynchronous (non-blocking) function calling, and proactive audio is always on.
Google publishes its Gemini Live evaluations in a model card and an evaluation PDF but does not list benchmark scores on its model pages. The Third-party figures in this section are published by Artificial Analysis, an independent benchmarking company, not by Google. Accessed October 16, 2026.
Realtime model results depend heavily on settings such as reasoning effort or thinking level: higher settings usually improve accuracy but increase time to first audio. Benchmarks from different providers are not directly comparable. Test candidate models with real calls from your own use case.