{
  "model_name": "gpt-live-1",
  "model_organization": "OpenAI",
  "submitting_organization": "OpenAI",
  "submission_date": "2026-09-16",
  "submission_type": "custom",
  "modality": "voice",
  "contact_info": {
    "email": "jiayuhuang@openai.com",
    "name": "Jiayu Huang",
    "github": "jiayuhuang-openai"
  },
  "results": {
    "banking_knowledge": {
      "pass_1": 37.11340206185567,
      "retrieval_config": "alltools"
    }
  },
  "is_new": true,
  "trajectories_available": true,
  "trajectory_files": {
    "banking_knowledge": "banking-high-elevenlabs"
  },
  "methodology": {
    "evaluation_date": "2026-09-16",
    "tau2_bench_version": "1.0.1",
    "user_simulator": "gpt-5.5 (xhigh)",
    "notes": "Custom voice evaluation on all 97 banking_knowledge base tasks, regular speech, alltools retrieval, one trial. Run seed 300; trial 0 seed 626729. Agent openai_live:gpt-live-1 delegates to gpt-6-astra at high effort, with marin voice and priority service. Caller openai/responses/gpt-5.5 uses xhigh effort and ElevenLabs eleven_v3 with native persona voices; full reviewer openai/gpt-5.5. Native error and hallucination retry limits are three; one 15-worker controller. Sierra tau2-bench 1.0.1, base 2174a603f6d014ef94473ffa95957f6ce27100db, with local execution/resource changes; task definitions, policies, prompts and native grading logic unchanged. The entire high run was selected from two complete ElevenLabs candidates (high and xhigh), with no per-task best-of mixing. Includes all 97 canonical histories, combined both.wav recordings, native label files, and original rewards and reviews. Differs from the 2026-09-09 GPT Live custom entry (medium backend effort, default Anthropic hallucination reviewer, 31.96% banking) in backend effort, service tier, and hallucination reviewer; task definitions, seed, voices and personas are identical. Sierra verification: checksums verified; 36/97 recounted; re-graded on main with embedded and fresh task definitions (0 flips); interaction_metrics recomputed (exact match). The run's results.json recorded audio_native_config.live_config.service_tier='priority', a field not in the public tau2-bench LiveConfig schema; it was removed from the hosted results.json so the trajectories load with public tooling, and no other trajectory content was modified.",
    "verification": {
      "modified_prompts": false,
      "omitted_questions": false
    }
  },
  "voice_config": {
    "provider": "OpenAI",
    "model": "gpt-live-1",
    "tick_duration_seconds": 0.2,
    "max_steps_seconds": 1200.0,
    "user_tts_provider": "elevenlabs/eleven_v3"
  },
  "interaction_metrics": {
    "version": "1.0",
    "config": {
      "tick_duration_sec": 0.2,
      "no_yield_window_sec": 2.0,
      "backchannel_yield_window_sec": 1.0,
      "vocal_tic_yield_window_sec": 1.0,
      "non_directed_yield_window_sec": 1.0,
      "vocal_tic_response_window_sec": 2.0,
      "non_directed_response_window_sec": 2.0
    },
    "domains": {
      "banking_knowledge": {
        "response_latency_mean": 3.4426586102719052,
        "yield_latency_mean": 1.1193181818181819,
        "response_rate": 0.8738120380147836,
        "yield_rate": 0.25844346549192365,
        "agent_interruption_rate": 0.15469904963041184,
        "selectivity_backchannel": 0.8421052631578947,
        "selectivity_vocal_tic": 0.7045454545454546,
        "selectivity_non_directed": 0.6923076923076923,
        "counts": {
          "n_simulations": 97,
          "response_total": 1894,
          "yield_total": 681,
          "backchannel_total": 19,
          "vocal_tic_total": 264,
          "non_directed_total": 221,
          "agent_interrupts_count": 293
        }
      }
    },
    "overall": {
      "response_latency_mean": 3.4426586102719052,
      "yield_latency_mean": 1.1193181818181819,
      "response_rate": 0.8738120380147836,
      "yield_rate": 0.25844346549192365,
      "agent_interruption_rate": 0.15469904963041184,
      "selectivity_backchannel": 0.8421052631578947,
      "selectivity_vocal_tic": 0.7045454545454546,
      "selectivity_non_directed": 0.6923076923076923,
      "counts": {
        "n_simulations": 97,
        "response_total": 1894,
        "yield_total": 681,
        "backchannel_total": 19,
        "vocal_tic_total": 264,
        "non_directed_total": 221,
        "agent_interrupts_count": 293
      }
    }
  },
  "reasoning_effort": "Backend: GPT-6 Astra (high)",
  "references": [
    {
      "title": "OpenAI Live provider integration",
      "url": "https://github.com/sierra-research/tau2-bench/pull/523",
      "type": "github"
    },
    {
      "title": "OpenAI Live prompt templates and evaluation runner",
      "url": "https://github.com/sierra-research/tau2-bench/pull/527",
      "type": "github"
    }
  ],
  "model_release": {
    "release_date": "2026-09-10",
    "announcement_url": "https://openai.com/index/introducing-gpt-live-1-in-the-api/",
    "announcement_title": "Build more natural voice experiences with GPT-Live-1 in the API"
  }
}