Cekura S2S - Data Accuracy: leaderboard
Cekura Bench realtime speech-to-speech leaderboard: each voice model runs as the agent on 82 scripted phone scenarios (59 patients booking, moving or cancelling clinic visits; 23 Medicare callers whose eligibility is checked and who are routed), three times each, inside the same open-source Pipecat agent with the same prompt and tools, over live WebRTC audio including noisy and glitchy calls. Labels carry the model setting Cekura states. Score: data accuracy (%), the share of calls in which every detail the agent saved matched the expected record; calls that save nothing are left out. Higher is better.
Metric: Data accuracy (%): calls whose saved details all matched. Source: benchmarks.cekura.ai. Saturation forecast: Rough model projection: around 2026. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Phonic 1 (Intelligence high) | 97.5 | |
| 2 | Cekura Cascade (flux-general-en → gpt-4.1 → eleven_flash_v2_5) | 94.1 | |
| 3 | GPT Realtime 2.1 (Reasoning high) | 92 | |
| 4 | GPT-Live 1 + GPT-6 Sol (Sol on low thinking) | 89 | |
| 5 | Grok Voice 3.0 (Reasoning high) | 87.3 | |
| 6 | Azure Realtime | 86.1 | |
| 7 | Gemini 3.8 Live Extended Thinking (Thinking level high) | 84.4 | |
| 8 | Gemini 3.1 Flash Live Preview (Thinking level high) | 79.2 | |
| 9 | Grok Voice Think Fast 2.0 (Reasoning high) | 78.9 | |
| 10 | Deepslate Opal | 72.6 |
Interactive version: theaggregate.ai/benchmark?slug=cekura-s2s-data-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-05.