A Model That Decides Many Times a Second
OpenAI shipped GPT-Live on July 8, 2026, built on a full-duplex architecture that generates its own speech output while it is still listening to the user. Instead of trading turns in sequence, it processes input and output continuously in parallel, deciding many times per second whether to speak, keep listening, pause, interrupt, or invoke a tool. Two tiers — GPT-Live-1 and the lighter GPT-Live-1 mini — are rolling out as the new default for ChatGPT voice mode, and developers reach the same models through the Realtime API.
The New Failure Point Background Delegation Creates
When a question needs search, deeper reasoning, or a complex task, GPT-Live hands the work to GPT-5.5 in the background while keeping the conversation going, then folds the result back in once it's ready. A text chatbot's tool call can leave the screen idle while the user waits, but in a voice call, that same wait has to be filled with natural stalling speech or follow-up questions or the call feels dead. This delegation path is a failure point text-only agents never had to handle.
From Barge-In to Deployment: Running a Full-Duplex Voice Agent
Set voice-specific target metrics separately from your text-agent baselines. Aim for a barge-in recognition rate above 95% for genuine interruptions, and a false-positive rate under 5% for background noise or coughs misread as interruptions. Keep the p95 round-trip latency for GPT-5.5 delegation under 3 seconds — long enough to stay inside the window that stalling speech can plausibly cover.
Failures cluster into three patterns. First, false-positive barge-in: a cough or background noise gets misread as an interruption and the model cuts itself off mid-sentence, which reads as unnatural; miss a genuine interruption instead and the user feels ignored. Second, dead air during delegation: if the GPT-5.5 response is delayed and there's no filler speech queued, the user assumes the call dropped.
Third, tone drift. Early users flagged the model coming across as overly enthusiastic, and that kind of persona drift can creep further off-baseline over the course of a long conversation. The recovery strategy is to wire an abort condition into each of the three patterns ahead of time: below a barge-in confidence threshold, keep listening instead of interrupting; when the delegated response exceeds its time budget, switch automatically to a pre-scripted set of stalling lines; and when tone drift is detected, force-reset the session back to the default persona.
Pre-deployment scenario testing has to include overlapping multi-speaker audio, noise-laced interruptions, and simulated delegation delays. Log utterance start and end timestamps, the barge-in verdict and its confidence score, whether delegation fired and its round-trip time, and safety-filter flags — the fields you need later to trace which decision went wrong.
Because streamed voice output starts playing before a sentence is even finished, harmful-content filtering has to run on streaming chunks, not completed responses. Route any PII captured in call transcripts through automatic masking before storage, and if the chunk-level filter fails to reach a verdict in time, fall back to muting that session's audio output immediately rather than letting it play.
Aggregate barge-in false-positive and missed-interruption cases, plus delegation-timeout cases, by type every week to see which noise patterns or query types cluster the failures. Keep the change log for model version swaps separate from the log of barge-in threshold tuning, so that when the next generation of voice model ships, you can isolate which change caused a regression.
The Operating Baseline to Set Now
A full-duplex voice model can't just inherit a text agent's turn-based safeguards. Fix three numbers first — a barge-in recognition rate above 95%, a false-positive rate under 5%, and p95 delegation latency under 3 seconds — then add chunk-level streaming safety filters and a tone-drift reset path, and a system making judgment calls several times a second becomes something a human can actually verify and operate.