OpenAIGoogleAudio matchupReal-time voice dialog
GPT-Realtime 2.1vs.Gemini 3.8 Live
Which model is better?
GPT-Realtime 2.1 does better in independent tests.
Pick GPT-Realtime 2.1 for
Pick Gemini 3.8 Live for
Shared tests
Independent evaluators and leaderboards only, the newest shared snapshot per evaluator. With 95 % ranges a lead counts when they do not overlap; for hit rates, when it holds in a 95 % test. Without any uncertainty figure at most “slightly ahead”.
Speko
GPT-Realtime 2.1 slightly aheadGPT-Realtime 2.1 is clearly ahead in 0 of 7 subtests and slightly in 6, Gemini 3.8 Live in 0 clearly and slightly in 1.
Speko S2S booking before caller confirmation
lower is better, as of October 28, 2026
Speko S2S task-content latency p50
lower is better, as of October 28, 2026
Speko S2S first-sound latency p50
lower is better, as of October 28, 2026
Show all 7 subtests
| Subtest | GPT-Realtime 2.1 | Gemini 3.8 Live | Verdict |
|---|---|---|---|
| Speko S2S booking before caller confirmation (lower is better) | 0.0 % | 18.0 % | GPT-Realtime 2.1 slightly ahead |
| Speko S2S task-content latency p50 (lower is better) | 1,145 ms | 1,569 ms | GPT-Realtime 2.1 slightly ahead |
| Speko S2S first-sound latency p50 (lower is better) | 1,022 ms | 1,242 ms | GPT-Realtime 2.1 slightly ahead |
| Speko S2S confirmation-safe action success | 78.0 | 68.0 | GPT-Realtime 2.1 slightly ahead |
| Speko S2S overall call quality | 84.0 | 79.0 | GPT-Realtime 2.1 slightly ahead |
| Speko S2S scripted call completion | 100.0 % | 96.0 % | GPT-Realtime 2.1 slightly ahead |
| Speko S2S task capability mean | 86.0 | 89.0 | Gemini 3.8 Live slightly ahead |
Source · latest October 28, 2026
Artificial Analysis
GPT-Realtime 2.1 aheadGPT-Realtime 2.1 is clearly ahead in 1 of 5 subtests and slightly in 1, Gemini 3.8 Live in 0 clearly and slightly in 1. 2 are level.
Artificial Analysis Big Bench Audio speech reasoning
1,000 tasks, as of October 15, 2026
Artificial Analysis τ-Voice task completion
as of October 15, 2026
Artificial Analysis Speech Agent Arena preference Elo
as of October 28, 2026
Show all 5 subtests
| Subtest | GPT-Realtime 2.1 | Gemini 3.8 Live | Verdict |
|---|---|---|---|
| Artificial Analysis Big Bench Audio speech reasoning | 96.0 % | 92.0 % | GPT-Realtime 2.1 ahead |
| Artificial Analysis τ-Voice task completion | 45.7 % | 30.1 % | GPT-Realtime 2.1 slightly ahead |
| Artificial Analysis Speech Agent Arena preference Elo | 937 | 1,079 | Gemini 3.8 Live slightly ahead |
| Speech Agent Arena task success rate | 91.5 % | 93.2 % | level |
| Artificial Analysis Conversational Dynamics | 95.7 % | 96.1 % | level |
Source · latest October 28, 2026
Cekura
GPT-Realtime 2.1 slightly aheadGPT-Realtime 2.1 is clearly ahead in 0 of 3 subtests and slightly in 3, Gemini 3.8 Live in none.
Cekura TurnBench end-of-turn score, Light noise
36 tasks, as of October 23, 2026
Cekura TurnBench end-of-turn score, Loud noise
36 tasks, as of October 23, 2026
Cekura TurnBench end-of-turn score, Echoey rooms
36 tasks, as of October 23, 2026
Source · latest October 23, 2026
All data side by side
Prices are verbatim from the provider, because audio models bill per minute, character or token.
| GPT-Realtime 2.1 | Gemini 3.8 Live | |
|---|---|---|
| Task | Real-time voice dialog | Real-time voice dialog |
| Developer | OpenAI | |
| Status | Active | Active |
| Weights | proprietary | proprietary |
| Runs locally | no | no |
| Price | $4 input, $0.40 cached input, $24 output per 1M tokens | $0.75 per 1M text tokens |
| Popularity (out of 100) | 41 | 0 |