OpenAI GPT-Live-1 API Launch - Full Duplex Bench v3 Tool calling: leaderboard
OpenAI's September 2026 GPT-Live-1 API launch evaluation on Full Duplex Bench v 3: spoken tool requests contain natural pauses, hesitations and self-corrections. Pass@1 scores the tool-call sequence, in percent; higher is better. GPT-Live-1 delegates to the reported Terra backend at low effort; the backend generation and Realtime comparators' effort are not specified.
Metric: Tool-call sequence pass@1 (%). Source: openai.com. Saturation forecast: Rough model projection: around 2026. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-Live-1 (Terra, Low) | 87 | |
| 2 | GPT Realtime 2.1 (API launch, effort unspecified) | 60 | |
| 3 | GPT Realtime 2 (API launch, effort unspecified) | 58 |
Interactive version: theaggregate.ai/benchmark?slug=openai-gpt-live-1-api-launch-full-duplex-bench-v3-tool-calling · How It Works · Data refreshed daily, snapshot 2026-11-05.