OpenAI published the engineering deep-dive behind GPT-Live: a full-duplex voice architecture that processes input and generates output at the same time, instead of waiting for a pause to decide whose turn it is to talk. It is a genuine architecture shift away from turn-based record-transcribe-respond voice bots, and it is worth understanding even for a business building on Claude, because who talks when is one of the hardest unsolved problems in voice AI.
What OpenAI built
The GPT-Live architecture centres on three design choices:
A model that listens and speaks at the same time, deciding whether to speak, keep listening, pause, interrupt, or invoke a tool, many times per second.
Lightweight acknowledgment sounds so the caller knows they are heard mid-sentence, rather than sitting in dead air.
A background handoff to a stronger reasoning model for anything that needs deeper thinking, without breaking the conversational flow the caller hears.
That is a real technical achievement. Full-duplex turn-taking, where a system genuinely listens and speaks concurrently rather than faking it with fast interruption detection, has been one of the harder open problems in conversational AI for years. It is the difference between a voice agent that feels like talking to dead air with a fast responder on the other end, and one that feels like an actual conversation.
Where Claude's voice mode places its bet
Claude's own voice mode, expanded in July across Opus, Sonnet and Haiku, takes a different bet. It leans on tool-calling into connected apps like Gmail and Slack rather than chasing full-duplex turn-taking as the headline feature. For most AU SMB use cases, a reception line, an intake agent, an internal ops assistant, the harder problem usually is not whether the system can interrupt naturally. It is whether it can actually do the task once it understands the request: book the appointment, pull the invoice, update the record, escalate to a human when it should.
If you are scoping a voice agent build, the practical question is not which lab has the smoother interruption model. It is which platform lets the voice layer safely trigger real actions in your systems, with the same audit trail you would expect from any other business automation, and the same admin-level controls over what it is allowed to touch.
Where full-duplex still matters
None of this means turn-taking is irrelevant. A high-volume consumer support line, or anything where callers routinely talk over the system, will feel the difference between full-duplex and a well-tuned interruption model. The point is narrower: for most Australian small business use cases we scope, the system's ability to safely execute a task matters more to the caller's experience than shaving another 200 milliseconds off the handoff between listening and speaking.
A Melbourne trades business fielding fifty booking calls a day, for example, gets far more value from a voice agent that can actually check trade availability and confirm a time slot in the job-management system than from one that interrupts marginally more naturally. The interruption model is the part a caller notices in the first three seconds; the task-completion rate is the part that determines whether the business keeps paying for the deployment past month three.
What to actually ask a voice-AI vendor
Whichever underlying model ends up in a voice agent, the evaluation checklist for an AU business should look past the demo and toward what happens after the call connects to a real system:
Can the voice layer read and write to your actual booking system, CRM or invoicing tool, or does it only summarise the call for a human to action later?
What happens when the caller asks for something outside its permitted scope, does it escalate cleanly to a person or improvise?
Is there a call transcript and action log an admin can review, the same audit trail you would expect from any other automation touching customer data?
Does the vendor's Australian Privacy Principles position hold up for recorded calls, given voice data is more sensitive than a typed chat log?
Most vendor demos are built to showcase the interruption model, because that is the part that sounds impressive in thirty seconds. The parts that determine whether the deployment survives its first month in production, the systems integration, the escalation paths, and the audit trail, rarely make the demo reel.
The Automata AI take
Expect the voice-AI architecture race between the major labs to keep making headlines through the rest of the year. For an Australian business actually deciding whether to build a voice agent, the build itself matters more than the underlying interruption model. That is the build Automata AI scopes for AU businesses, typically A$3,500 to A$8,000 depending on how many tools it needs to connect to and how much auditing the compliance side requires.
Talk to us about a voice agent build and we will map out what a Claude-based voice layer would actually need to touch in your systems.
