The GPT-Live-1 interruption test that tells you whether to retire your STT-LLM-TTS chain

Your voice agent just got a new rival: OpenAI made GPT-Live-1 available through its API on September 10, 2026. Language-learning company Speak reported that GPT-Live-1 produced almost 80% fewer interruptions during learner thinking pauses than its previous turn-based systems. If you run telephony, support, or conversational-agent workflows, the question is whether your STT-LLM-TTS chain still wins on the calls you actually take.

A cascade is a relay where every handoff costs time. A full-duplex model tries to remove the handoff. The tradeoff is the whole decision.

The replacement test is a scorecard, not a demo

Treat the cascade as the baseline and GPT-Live-1 as the challenger. Score both on the same five points. A vendor demo can hide the exact pauses, tool calls, and long calls that affect your queue.

• Interruption rate. Count how often the agent cuts off a caller during thinking pauses, and compare that against your current stack.
• Long-session retention. Track whether the agent still holds context, task state, and caller intent after a long call.
• Delegation latency. Measure the delay from caller request to an early useful response, including any handoff to a human or backend tool.
• All-in cost per minute. Add voice, model, tool, telephony, and engineering maintenance costs for the same traffic.
• Fallback failure mode. Decide what happens when the model is wrong, unavailable, or stuck, and whether the caller can recover.

Interruption and latency are the felt criteria

Full-duplex voice sounds like a small upgrade until a caller pauses to think. A turn-based stack treats that pause as silence and keeps talking. A full-duplex model has to decide whether the pause is a hesitation, a backchannel, or a real interruption. The decision is where your support queue feels the difference.

OpenAI said GPT-Live-1 outperformed GPT-Realtime-2.1 by 30 percentage points on Full Duplex Bench and placed at the top of Tau3 while using GPT-6 Astra with medium reasoning effort. The benchmarks are not your call logs. Use your own pauses: caller hesitation, background noise, agent mid-sentence, and the moment a caller says wait.

Turn-taking latency is the number a caller feels as a dead air gap. OpenAI's published comparison shows GPT-Live-1 turn-taking latency of about 0.8 seconds, compared with well over a second for predecessor voice models. The gap can feel smooth in a controlled test and rough in a call center where the caller is already impatient. Measure the same path: caller speech to agent response, including network, telephony, and any backend tool call.

Cost, retention, and fallback decide whether the win survives

GPT-Live-1 voice sessions cost $0.05 per minute with per-second billing, and backend model and tool usage is billed separately. The line item is easy to miss when a call triggers a lookup, a ticket update, or a handoff. The cascade has its own line items: speech recognition, model tokens, speech synthesis, telephony, and the code that keeps them in sync.

One customer said that moving from a cascaded voice system to GPT-Live-1 reduced its voice codebase by 80% and eliminated 23,000 lines used for real-time patient conversations. Maintenance savings like this can matter more than the per-minute price. GPT-Live-1 launched with 12 voices covering accents, dialects, and languages, and custom voice access requires contacting sales. If your brand voice or regional accent needs a custom model, that is a procurement item, not a toggle.

A voice agent that handles a short exchange well can still fail on a long support call. Track whether it keeps the caller's issue, prior answers, and task state across a long session. If it loses the thread, the caller repeats themselves, and the cost is not just minutes.

Fallback is where the controlled test ends. If the model is wrong, unavailable, or stuck, the caller needs a path to a human, a retry, or a safe exit. A cascade can be boringly reliable here because each stage has a known failure mode. A full-duplex model can be smoother, but you still need a clear stop.

Retire the chain only when your numbers say so

Put the five checks in a spreadsheet before you pilot. If the numbers favor GPT-Live-1, run a staged rollout. If the cascade still wins, fix the weakest component.