Speech Agent Arena LeaderboardArtificial Analysis
Compare Speech to Speech models by preference across blind, live voice conversations on topics like booking a dental appointment and ordering takeout. For more details, see overview and methodology, or visit Compare Models & Providers to compare Speech to Speech models across other benchmarks.
Range | Model | Elo | Samples | Task Success Rate | TSR 95% CI | |
|---|---|---|---|---|---|---|
| 1 | 1-2 |  Gemini 3.1 Flash Live Minimal | 1095±2695% interval 1069 to 1121 | 935 | 74.6% | 69.4%/79.2% |
| 2 | 1-4 | Gemini 3.8 Live | 1078±2895% interval 1050 to 1105 | 557 | 93.2% | 89.0%/95.8% |
| 3 | 2-6 | Gemini 3.1 Flash Live High | 1066±2595% interval 1041 to 1092 | 913 | 71.8% | 66.4%/76.7% |
| 4 | 2-6 |  Grok Voice 3.0 | 1064±2895% interval 1036 to 1093 | 637 | 94.8% | 91.6%/96.8% |
| 5 | 3-7 |  GPT-Live-1 (Sol, low) | 1047±2695% interval 1021 to 1073 | 730 | 90.9% | 87.2%/93.7% |
| 6 | 3-7 | GPT-Live-1 (Astra, medium) | 1044±2695% interval 1018 to 1070 | 681 | 87.4% | 82.9%/90.8% |
| 7 | 5-11 |  Cartesia Line (Default Cascaded System) | 1024±3395% interval 990 to 1057 | 266 | 77.5% | 68.9%/84.3% |
| 8 | 8 | GPT-Realtime-1.5 | 1000AnchorFixed at 1000 as the reference point for this board's Elo scale. | 876 | 85.1% | 80.8%/88.6% |
| 9 | 7-14 |  Eleven Realtime v1 Max | 997±2895% interval 970 to 1025 | 638 | 87.1% | 82.6%/90.6% |
| 10 | 8-14 | Grok Voice 2.0 | 997±2695% interval 971 to 1023 | 624 | 94.6% | 91.4%/96.7% |
| 11 | 8-14 | ElevenLabs Agents (Default Cascaded System) | 994±2595% interval 969 to 1019 | 840 | 90.5% | 86.7%/93.4% |
| 12 | 8-15 | Eleven Realtime v1 Mini | 988±2895% interval 960 to 1016 | 613 | 60.3% | 54.3%/66.0% |
| 13 | 8-15 | Gemini 3.8 Live Extended Thinking (High) | 979±2595% interval 954 to 1004 | 724 | 89.1% | 84.7%/92.4% |
| 14 | 9-15 |  Nova 2.0 Sonic (Mar 2026) | 973±2495% interval 949 to 998 | 869 | 57.1% | 51.1%/62.9% |
| 15 | 12-15 | GPT Realtime (Aug '25) | 969±2595% interval 944 to 993 | 862 | 89.4% | 85.2%/92.6% |
| 16 | 16-21 | GPT-Realtime-2.1 High | 937±2595% interval 912 to 962 | 835 | 91.5% | 87.7%/94.3% |
| 17 | 16-21 | GPT-Realtime-2.1 Minimal | 937±2595% interval 912 to 961 | 874 | 89.4% | 85.2%/92.5% |
| 18 | 16-23 | GPT-Realtime-2 (High) | 926±2595% interval 901 to 950 | 855 | 89.8% | 85.7%/92.9% |
| 19 | 16-23 | GPT Realtime Mini (Oct '25) | 924±2495% interval 900 to 948 | 909 | 79.6% | 74.6%/83.8% |
| 20 | 16-23 | GPT-Realtime-2 (Minimal) | 924±2495% interval 899 to 948 | 922 | 84.7% | 80.3%/88.3% |
| 21 | 16-23 | Nova Sonic | 913±2695% interval 887 to 938 | 827 | 57.1% | 51.6%/62.5% |
| 22 | 18-23 | Grok Voice 1.0 | 908±2695% interval 882 to 934 | 864 | 80.7% | 75.8%/84.8% |
| 23 | 18-23 |  Deepgram Voice Agent (Default Cascaded System) | 908±2595% interval 883 to 932 | 886 | 73.7% | 68.5%/78.2% |
| 24 | 24 |  Inworld Realtime (Default Cascaded System) | 882±2595% interval 857 to 907 | 880 | 69.9% | 64.3%/75.0% |
| 25 | 25-28 |  Qwen3.5 Omni Plus Realtime | 852±2595% interval 827 to 878 | 888 | 62.0% | 56.4%/67.2% |
| 26 | 24-28 | Qwen Audio 3.0 Realtime Flash | 847±3695% interval 811 to 883 | 245 | 81.7% | 70.1%/89.4% |
| 27 | 25-28 | Qwen3.5 Omni Flash Realtime | 843±2595% interval 818 to 868 | 922 | 29.1% | 24.0%/34.8% |
| 28 | 25-28 | GPT-Realtime-2.1 Mini Minimal | 840±2595% interval 814 to 865 | 912 | 76.7% | 71.1%/81.4% |
| 29 | 29 | Qwen Audio 3.0 Realtime Plus | 793±3995% interval 755 to 832 | 267 | 77.8% | 66.9%/85.8% |