All tasks / Comparisons
Gemini 3.8 Live vs GPT-Realtime-2.1
Compared only on public benchmark results, across the 1 task where both are listed: Gemini 3.8 Live is a tier higher in 1 of 1 shared task.
Gemini 3.8 Live
Google's native-audio Gemini models for low-latency voice agents in the Live API, with an Extended Thinking variant.
- Vendor
- Weights
- Closed: API and apps only
- Released
- Sep 2026
- API price
- $0.01 per minute of audio (Artificial Analysis 'Cost per Hour of Input Audio' for Gemini 3.8 Live, Google ($0.841/hour: cost of its Big Bench Audio run normalised to input-audio length) divided by 60.)
- License
- proprietary
- Last verified
- 18 Oct 2026
GPT-Realtime-2.1
OpenAI's speech-to-speech models for the Realtime API, with configurable reasoning effort and tool use for voice agents.
- Vendor
- OpenAI
- Weights
- Closed: API and apps only
- Released
- 6 Jul 2026
- API price
- $32 in / $64 out per 1M tokens (Audio tokens, USD per 1M (text tokens $4 in / $24 out).); $0.19 per minute of audio (Artificial Analysis 'Cost per Hour of Input Audio' for GPT-Realtime-2.1 Minimal, OpenAI ($11.307/hour: cost of its Big Bench Audio run normalised to input-audio length) divided by 60.)
- License
- proprietary
- Last verified
- 18 Oct 2026
Task by task
| Task | Gemini 3.8 Live | GPT-Realtime-2.1 | Higher tier |
|---|---|---|---|
| Voice agents Models | S1,088 on Artificial Analysis Speech Agent Arena (models) · Cloud | B941 on Artificial Analysis Speech Agent Arena (models) · Cloud | Gemini 3.8 Live |
A tier spans a range of scores, and small gaps on a leaderboard are often within its margin of error, so only a different tier counts as ahead. How we rank.
Same benchmark, both results
| Benchmark | Gemini 3.8 Live | GPT-Realtime-2.1 | Snapshot |
|---|---|---|---|
| Artificial Analysis Big Bench Audio (speech reasoning) Accuracy on Big Bench Audio: spoken reasoning questions answered in audio (speech in, speech out), higher is better | 97.7% | 95.8% | 18 Oct 2026 |
| Artificial Analysis Speech Agent Arena (models) Elo from blind pairwise votes on live voice-agent conversations (overall slice; GPT-Realtime-1.5 anchored at 1000), higher is better | 1,088 | 941 | 18 Oct 2026 |
| Artificial Analysis Speech to Speech Index AA-Speech to Speech Index: weighted average of speech reasoning (Big Bench Audio), agentic performance (tau-Voice), arena preference and task success rate, higher is better | 83 | 75 | 18 Oct 2026 |
| Cekura Bench: voice agent workflow benchmark v1 (speech-to-speech models) Reliability pass^3: share of 82 phone-call scenarios a reference agent on the model passes on all three runs, higher is better | 30.5% | 64.6% | 12 Oct 2026 |