For the GPT-Live-1's intelligence, Tau3 tests whether a voice agent can actually complete customer-support tasks across airline, retail, and telecom scenarios.

Paired with GPT-6 Astra at medium reasoning effort, GPT-Live-1 completed 83.6% of tasks on the first attempt, compared with 45.7% for GPT-Realtime-2.1.

The same pairing scored 38.1% on TauBanking, where agents need to find information in banking documents and apply it while using tools to resolve customer requests.

Of course, people chatting with our builds don’t talk in perfectly timed turns.

They pause to think, interrupt with another question, or say “mm-hmm” without expecting you to stop:
• GPT-Live-1 scored 97.3% on Artificial Analysis’s Conversational Dynamics benchmark, which evaluates when to wait, take a turn, respond to an interruption, or keep speaking through a short acknowledgment.
• On Full Duplex Bench v1.5, it scored 80.1% on interactivity 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. That measures the wait before the reply starts, not how quickly the model speaks.

Getting the task right also means following a request when someone pauses, hesitates, or corrects themselves.

With a backend model handling tool calls, the GPT-Live-1 setup achieved these results on Full Duplex Bench v3:
• 87% tool-calling success for completing the correct sequence of tool calls.
• 90% on the separate response-quality evaluation.

Read more in the launch blog:
https://t.co/cCtLbRRyLS