GPT-Live-1 review: is OpenAI's full-duplex voice model worth it?

Why I am reviewing this, and where I am coming from

I build product features and AI agents at eesel, and my team has spent the last three-plus years putting AI agents on live support queues. That vantage point shapes how I review a voice model. I have watched a confident-sounding bot quietly hand a customer the wrong answer, which is why we now simulate every rollout against a company's historical tickets before it ever talks to a real person. So when I test something like GPT-Live-1, I am not just asking "does it sound good," I am asking "what happens when it is wrong, and who is paying for the parts that make it right."

This is a hands-on review of the model itself: what it feels like to use, how it actually works under the hood, where it shines, where it stumbles, and what it means if you are thinking about voice for conversational support. I will lean on OpenAI's own launch materials for the mechanism and on real user reports for the reception.

What GPT-Live-1 actually is

GPT-Live-1 is OpenAI's third-generation voice model, first shipped inside ChatGPT Voice in July 2026 and released in the API on September 10, 2026. The headline feature is a full-duplex architecture: the model can listen and speak at the same time, rather than waiting for you to finish, guessing that you are done, and only then starting to think.

That sounds like a small thing. It is not. Older voice systems relied on a tiny "turn detector" model to decide when you had stopped talking. Guess too early and it cuts you off; guess too late and it feels sluggish. GPT-Live removes that detector from the audio path entirely, so the back-and-forth lands closer to human timing. In practice, it will drop in a "mhmm" while you are still talking, stay quiet when you pause to think, and let you interrupt it mid-sentence without losing the thread.

It replaces the old cascaded pipeline, the one that ran speech-to-text, then a model, then text-to-speech as a chained voice stack in series. If you have ever priced a voice stack by pairing something like ElevenLabs or Cartesia Sonic against a separate reasoning model, GPT-Live-1 collapses most of that into one model. One healthcare developer, Tony Stoyanov, said moving to it let his team cut 23,000 lines of code and simplify their codebase by about 80%.

How it works under the hood

This is the part I find really clever, and it is also the key to understanding the pricing and the limits. GPT-Live-1 puts the voice model in charge of the conversation and pushes everything else off the live path.

When a question needs web search, real reasoning, or a tool call, the voice model delegates to a backend frontier model (GPT-5.5 at launch) on a separate asynchronous path, and keeps talking to you while that work runs in the background. OpenAI rebuilt the media frontend in Go to replace an older Python implementation, using WebRTC for transport, and reports that the new system's p95 latency matches the old system's p50. The result is a clean split: a small, fast voice loop that must never stall, and a slower "brain" that can take its time without freezing the conversation.

That architecture is why your real cost is a stack, not a single number. On the API you are paying for three layers.

1. The voice layer: the flat $0.05/min for GPT-Live-1 itself.
2. The backend model tokens: whatever reasoning model you route to, billed separately and varying with how hard it thinks.
3. The agent harness and tools: your framework, retrieval, and any ticketing system writes.

So the honest read is that $0.05/min is the floor, not the bill. If you want the full math on that, we broke it down in the companion OpenAI API cost post.

What is really good

Let me give credit where it is due, because the wins here are real.

The naturalness is a step change. OpenAI reports GPT-Live scoring roughly 30 points higher on its Full Duplex Bench than the previous GPT-Realtime-2.1 model, and in blind testing people preferred it about 75.7% of the time over the older Advanced Voice Mode. Partners building on it, like the language app Speak, reported cutting unwanted interruptions by around 80%. Those numbers match my own impression: it is the first voice model where I stopped consciously managing my turns.

It is smart enough when paired well. With a medium-effort backend model, OpenAI puts GPT-Live at the top of its Tau3 agentic benchmark. The delegation trick means you get low-latency chat and frontier-model reasoning, as long as you wire up the backend.

On the consumer side it is effortless. In ChatGPT there is no per-minute meter at all. GPT-Live-1 is the default voice model on Go, Plus, and Pro, GPT-Live-1 mini is the default on Free, and it rolled out globally across iOS, Android, and the web on day one.

The model can also surface visual answers during a spoken conversation, like a map or a schedule, which makes the "talk to it while driving or cooking" use case land better than earlier voice assistants managed.

Where it stumbles

A fair review names the limits plainly, and there are a few worth knowing before you commit.

It is not the frontier model, and it shows. Because GPT-Live-1 is a voice layer that hands off reasoning, its own answers can be shallow or occasionally odd when it does not delegate. One user on Hacker News put it bluntly:

"GPT-Live isn't a particularly intelligent model. It needs to handoff for reasoning and more complex tasks. You shouldn't expect its responses to be as reliable as a frontier model. It will sometimes say odd things."

kypro, Hacker News

The conversational habits can grate. The same reflexes that make it feel human can tip into over-eager. A common complaint is that it interrupts thinking pauses and paraphrases you too much:

"The new gpt-live feature tends to interrupt with 'human' sounding injections like 'hmm', 'huh' a bit too much for my liking. Also its crutch is to start by paraphrasing what I just said to it... after every single thing I say it starts to feel a bit much."

kingkongjaffa, Hacker News

That said, plenty of real-world use is positive. One reviewer described leaning on it through CarPlay to practice French, noting it handled switching between languages mid-response "surprisingly well." The reception is mixed in the way brand-new interfaces usually are, not uniformly negative.

Custom voices and enterprise features are gated. Building your own branded voice means going through a "contact sales" flow, and OpenAI's enterprise voice product, Presence, sits as a separate layer on top. The self-serve API gives you the model, not the whole product.

My verdict, scored

Here is how I would score GPT-Live-1 across the dimensions I actually care about when evaluating a voice model for real work.

What GPT-Live-1 means for customer support

This is where I put my support hat back on, because voice-for-support is where the per-minute model gets interesting, and a little risky.

Per-minute billing charges you for talk time, not for outcomes. For a language tutor or a reservation bot, that is fine. For support it inverts the incentive you want. A frustrated customer on a ten-minute call that ends in an escalation costs you more than a calm one resolved in ninety seconds, even though the second interaction is the actual win. If first-contact resolution is what you report on, a meter that runs on minutes is measuring the wrong thing and quietly distorts your support metrics.

There is also the intelligence gap. GPT-Live-1 makes the conversation feel human. It does not, on its own, know your refund policy, resolve a ticket, or handle tier-1 deflection against your own docs. That is not a knock on OpenAI, it is by design: they built a voice layer and priced it as one. The business context and the automated ticket resolution are meant to come from whatever you build on top. This is the same lesson I see with any raw AI call center build: the voice is the easy, impressive part, and the knowledge, retrieval, and escalation handling are the parts that actually close the loop.

Where eesel fits: infrastructure versus the teammate

The clean way to frame it is that GPT-Live-1 is infrastructure, and eesel is the employee. GPT-Live-1 is a model you wire into your own product. eesel is a platform where you hire a ready-to-work AI teammate for a specific job, and it arrives already carrying the skills, integrations, and company context for that role.

For support, that teammate is the AI helpdesk teammate. It joins your existing queue, learns from your past tickets and help center, and, unlike a raw voice model, it is billed per resolution rather than per minute of talk, so your cost tracks tickets actually closed. You can still point it at a voice front end like GPT-Live-1 when you want natural phone conversations, and keep the cost-per-resolution math sane. And because we simulate on your historical tickets before go-live, you see the likely outcome before a customer ever does, which is the exact "what happens when it is wrong" question I opened with.

Because a lot of teams reaching for a voice API really want programmatic, headless control, it is worth knowing eesel exposes a real CLI alongside the dashboard. It is the same AI teammate and the same workspace, driven from a terminal instead of a browser: a person can run it by hand, scripts can automate it, and coding agents like Claude Code, Codex, and Cursor can operate it directly, the same way you would drive any other ChatGPT agents workflow. So if the reason you were eyeing the API was scriptable control, you do not have to trade "programmable" for "already knows my support context."

If you want the honest comparison for your own volume, the cost savings breakdown and our build vs buy guide both walk the per-minute-versus-per-resolution decision, and the wider case for AI vs human support sits alongside it.

GPT-Live-1 review, in one line

GPT-Live-1 is the best voice layer OpenAI has shipped, and it is exactly that: a layer. As a conversational AI interface it is a clear buy, free inside paid ChatGPT and a genuine leap in naturalness. As the foundation for a voice product you have to remember it is half the stack, with a backend model and harness billed on top. And if the job is support, ask the only question that matters: are you paying for minutes, or for resolved tickets?

Want a support teammate that already knows your help center and bills per resolution, not per minute? eesel plugs into your existing helpdesk in minutes, simulates on your own past tickets before it goes live, and is free to try. Try eesel or book a demo to see it on your own data first.