Perplexity Puts GPT-Live-1 Behind Its Voice Mode: What the Mac and Web Switch Actually Changes

Perplexity's voice mode on the Mac app and on the web now runs on GPT-Live-1, OpenAI's full-duplex speech model, and the capability attached to that swap is the part worth reading twice: a single spoken session can drive several Perplexity Computer sessions at once, start new tasks mid-sentence, and create or manage Projects without touching a keyboard. The announcement went out from Perplexity's @AskPerplexity account on 28 October 2026 and applied to paid Pro and Max subscribers on both surfaces.

That is a distribution event, not a model launch. GPT-Live-1 has been generally available since OpenAI put it on the API on 10 September 2026, at $0.05 per voice minute, and it has been in ChatGPT Voice since 8 July. What changed this week is who is renting it, and what they are pointing it at.

Why a voice-model swap is worth a headline

Voice interfaces used to be assembled inside each company. You bought transcription, called your own language model, and bought synthesis, then spent the hard part of the build on turn detection — the rule that decides when the user has finished speaking. That rule is where most voice products die: too eager and the agent talks over people, too slow and every call feels like a bad international line.

GPT-Live-1 removes that seam. It is one model that listens while it generates, decides many times per second whether to speak or wait, and hands heavy reasoning to a text model in the background without pausing the audio. Perplexity did not have to build a barge-in policy to get multi-session control working; it inherited one.

The practical consequence for readers building their own voice products is blunt: the turn-taking layer you were going to write is now something you rent. Whether that is cheaper depends entirely on how many minutes your agent spends listening rather than talking — a question worth doing arithmetic on before you commit, because the meter runs on the whole session, silence included.

The Computer part is the real product change

Model swaps are invisible to users. Multi-session Computer control is not. Until this update, Perplexity Computer was a single agent loop you supervised on screen. The new voice layer lets you hold one conversation while several background sessions run, kick off a new one by saying so, and manage Projects by name. That is closer to a dispatcher than to a chatbot, and it is the first consumer voice interface we have seen that treats parallel agent sessions as the default rather than the exception.

Three details are doing the work here:

• One session, many workers — a voice turn can address more than one Computer session without ending the current one.

• Task creation by speech — new work can be queued mid-conversation instead of after it.

• Project management by name — Projects become objects you can create and rename out loud, which means the voice layer needs to resolve references, not just transcribe.

None of that requires a full-duplex model in theory. All of it is much easier with one, because the agent can acknowledge a request with "mhmm" while the actual work starts, instead of going silent for eight seconds and hoping the user waits.

Perplexity named v1. OpenAI has since shipped v1.1

The sharp detail in the announcement is the version string. Perplexity says its voice mode runs on GPT-Live-1. In ChatGPT Voice, OpenAI began rolling out GPT-Live-1.1 on 15 October 2026, two weeks before this switch. Perplexity has not said whether it evaluated 1.1, whether it is pinned deliberately to v1 for stability, or whether an upgrade is scheduled.

Read that as normal engineering rather than a scandal. A platform integrating a speech model into a consumer app cares about a frozen snapshot with predictable latency far more than about the newest checkpoint, and OpenAI has not published a migration guide that makes 1.1 a drop-in. But it does mean the model in the Perplexity app today is not the model in ChatGPT today, and any comparison of the two voice experiences is measuring different weights.

The economics Perplexity has not talked about

OpenAI's own documentation lists one price for this model: $0.05 per minute of voice, billed by the second and not rounded up, with a 15-second session initialization charge credited against later duration. Reasoning, tools and web search run on a separate text model and are billed separately. Perplexity sells a flat subscription.

Those two billing models do not naturally fit. A Max subscriber who uses voice for two hours a week — not an unusual amount for someone who just gained hands-free control of several agent sessions — consumes roughly 520 minutes a month. At list price that is about $26 of voice alone, before the reasoning calls behind each Computer task. A subscriber who leaves a session open while thinking, or who keeps it running as ambient control, costs more, because silence bills at the same rate as speech.

Every platform that resells a metered voice model under a flat plan is making this bet, and the bet is usually right at launch: most subscribers use far less than the heavy tail. The interesting number to watch is not the average but the 95th percentile, and neither company publishes it.

On OpenAI's side, the per-hour figure that matters to capacity planning is input audio cost, not talk time. Artificial Analysis measured GPT-Live-1 at $5.83 per hour of input audio on its high-reasoning configuration and $4.47 on its low one, using a fixed 40-question benchmarking set — a useful reminder that the $0.05 headline covers voice only.

How good is the model Perplexity just adopted?

On Artificial Analysis' independently run Speech-to-Speech Index, GPT-Live-1 scores 83.2 in its Astra-medium configuration, second of 21 scored models behind Grok Voice 3.0 at 86.9, and 80.1 in its Sol-low configuration at fifth. On the same lab's Speech Agent Arena — blind preference over real booking-style conversations — the two configurations sit at 1037 and 1041 Elo, with task success rates of 87.4% and 90.9%.

The configuration string is not decoration. GPT-Live-1 does not reason in-house; it delegates to a named backend, and which backend you pick changes both the score and the bill. Perplexity has not disclosed which backend it pairs with the voice layer, so its effective quality is unknown from the outside even though its voice model is now public knowledge.

The routing question nobody asked

Here is the structural point this announcement makes visible. Perplexity's voice leg is now rented from OpenAI and cannot be substituted — GPT-Live-1 exposes exactly one endpoint, v1/live/sessions, and is explicitly not a drop-in for the Realtime API. That leg is locked. The reasoning leg behind it is not.

Which is where a router earns its place: the text model doing the thinking for each Computer task can be swapped, failed over, or split by task type without touching the voice layer. On OrcaRouter, the same key reaches GPT-6 Astra, GPT-5.6 Terra and GPT-5.6 Sol — the tiers OpenAI itself uses as delegation targets — alongside Claude, Gemini, Qwen and GLM models, with provider list prices passed through at 0% markup and automatic failover when one backend degrades. When a vendor cuts a price, the routed leg reflects it the same day.

To be precise about what we do not do: OrcaRouter does not serve GPT-Live-1. The voice leg stays on OpenAI's API. What we route is everything downstream of it, which is where a voice agent's cost and reliability variance actually lives.

What to watch next

Three things would change this story. If Perplexity moves to GPT-Live-1.1, the version note above stops mattering and the app's voice quality should step up with whatever 1.1 changed. If OpenAI publishes per-session cost controls — a cap, a silence timeout, a cheaper idle rate — the flat-subscription math gets easier for every reseller. And if a rival full-duplex model lands on a comparable index, the single-vendor dependency in this integration becomes a choice rather than a default.

For developers, the actionable read is narrower. If you are building an interruption-heavy voice interface — intake, booking, dispatch, anything where the user talks more than the agent — a full-duplex model is now the sane default and per-minute billing is the price of it. If you are building narration, read-aloud, or anything that produces audio from text on demand, this model is priced and shaped for the wrong job, and the TTS comparison is a different article entirely.