OpenAI Launches GPT-Live-1 Voice Model, Achieving 83.6% First-Attempt Task Completion in Customer Support Benchmark

OpenAI unveiled GPT-Live-1, a new voice model now available through its API, on September 11, 2026. In benchmark results shared by the company, GPT-Live-1 paired with GPT-6 Astra at medium reasoning effort completed 83.6% of customer-support tasks on the first attempt in the Tau3 benchmark, compared with 45.7% for GPT-Realtime-2.1.

Tau3 evaluates whether a voice agent can complete customer-support tasks across airline, retail, and telecom scenarios. The same pairing scored 38.1% on TauBanking, where agents must locate information in banking documents and apply it while using tools to resolve customer requests.

On conversational benchmarks, GPT-Live-1 scored 97.3% on Artificial Analysis's Conversational Dynamics benchmark, which measures when to wait, take a turn, respond to an interruption, or keep speaking through a short acknowledgment. It scored 80.1% on interactivity in Full Duplex Bench v1.5 across overlapping speech, background voices, side conversations, and interruptions. In Full Duplex Bench v1, it began responding 0.798 seconds after the user's turn ended, compared with 1.41 seconds for GPT-Realtime-2.1.

In test setups where a backend model handles tool calls, the GPT-Live-1 configuration achieved 87% tool-calling success on Full Duplex Bench v3 and 90% on the separate response-quality evaluation. OpenAI published further details in a launch blog post.