GPT-Live-1 vs Sesame Preview: The Best Voice in This Comparison Is the One You Cannot Buy

Ask anyone who has used both and the ranking is not close: Sesame Preview is the most convincing voice assistant most people have spoken to. It is also the one you cannot build on. GPT-Live-1 and Sesame Preview are compared constantly — both are conversational, both handle interruption, both sound like a person rather than a phone tree — but they are offered in opposite ways. GPT-Live-1 is a hosted model you rent by the minute, publicly callable since September 10, 2026. Sesame Preview is a free consumer app with no API at all, and the voice people are impressed by runs on a production model that has never been released.

What each one actually is

• GPT-Live-1 — Ope​nAI's full-duplex voice model, in ChatGPT since July 8, 2026, in the API since September 10 at $0.05 per voice minute. Twelve API voices, no cloning, no image or video input, transcripts and response text returned with each session.

• Sesame Preview — Sesame's consumer voice assistant. Free on iOS since May 2026, broadly available on Android since early August 2026, plus a voice-first web preview. Four agents — Maya, Miles, Simone and Charlie — English only, with a 30-minute call cap that drops to 5 minutes when logged out.

• CSM-1B — the part of Sesame that is open: a 1B conversational speech model released under Apache 2.0 on April 23, 2026, built on a Llama-architecture backbone with a separate decoder and Kyutai's Mimi codec, producing 24 kHz mono English speech from a context of roughly 4K tokens.

The three lines above are the whole shape of the decision. One company sells a model by the minute. The other gives away an app and open-sources a smaller model that is not the one in the app.

The gap between the demo and the download

Sesame has been unusually direct about this, and it is the single most important fact for anyone evaluating the pair. The voice in the Preview app — the one that produced the viral clips and the "best-in-class voice mode" reviews — is a larger, closed production model. CSM-1B is a 1B-parameter open checkpoint from the same lab, and by the assessment of people who have run both, it is a distinctly weaker voice. Sessions on Preview are also capped and limited to English, and there is no task execution: the agents talk, they do not book, buy, or file anything.

That leaves developers with a real but narrower offer: a self-hostable conversational speech checkpoint with no licence fee, hosted by one third-party inference provider at $7 per million characters — roughly $0.35 per hour of synthesised audio — or free on your own hardware. Nobody has published an independent benchmark of either the Preview voice or CSM-1B, so any quality claim in either direction is a listening impression, not a measurement. Sesame has also published no public pricing, no enterprise tier and no monetisation plan; the company's stated direction is research partnerships and a planned hardware product.

What $0.05 a minute buys that free does not

GPT-Live-1's pitch to a developer is not that it sounds better, because that argument is unwinnable against a closed model nobody can measure. It is that the thing is callable and priced. Concretely, what you get for the meter running:

• A session you control — one model ID, one Live Sessions endpoint, the published integration example running over WebRTC, and a session you configure with instructions, voices and a delegation block.

• Interruption handling as a documented behaviour — 80.1% interactivity on Full Duplex Bench v1.5 against 45.4% for GPT-Realtime-2.1, with 0.798-second turn-taking latency and 87% tool-calling success. All three are Ope​nAI's own numbers, published with the release and unreproduced by anyone at the time of writing.

• A reasoning backend you choose — the voice layer hands heavy work across an asynchronous boundary to a separate model named in the session configuration, which keeps working while the conversation continues. In Ope​nAI's documentation example that backend is GPT-5.6 Terra; at the July consumer launch it was GPT-5.5.

• Transcripts, response text and telephony-shaped billing — $3.00 for an hour of continuous open line, charged by the second, with 15 seconds of session initialisation credited rather than added.

Sesame gives you a better conversation and none of those. If you are building a product, "better conversation, no API, no price, no benchmark, English only, 30-minute cap" is not a comparison — it is a waitlist.

There is a number on GPT-Live-1 now that did not exist at its consumer launch, and it is the honest counterweight to everything above. Artificial Analysis's Speech to Speech Index scores GPT-Live-1 at 69.8% — seventh of eleven models, behind GPT-Realtime-2.1 at 73.9% and behind Grok Voice Think Fast 2.0 at 79.0%. So the model you can buy is not the best one on the independent board either. What the board cannot score is the thing Sesame is actually winning on: none of the eleven models on that index was rated on how it feels to talk to, and the Sesame Preview voice has no row anywhere.

The piece of the stack that is neither company's

Here is where both options converge, and where the decision stops being binary. Neither Sesame Preview nor GPT-Live-1 is a complete agent. Sesame's agents do not execute tasks; GPT-Live-1 explicitly delegates anything that requires reasoning, retrieval or a tool to a backend model. In both designs the interesting work happens behind the voice, in a plain text model call, and that layer is portable in a way the voice layer is not — you can move a backend without re-recording a single prompt for the speech model.

That backend is also where OrcaRouter is useful, and it is worth being precise about what we do and do not carry. We do not route GPT-Live-1: the Live Sessions endpoint is Ope​nAI's, and the voice layer there is out of reach. We do not route Sesame's production model either, for the simpler reason that it has never been offered as an API. What we do route is the layer both products hand work to. Roughly 190 models from eleven upstream providers run behind one key at provider list price with no markup, with automatic failover across providers and a routing DSL that can compose several models into a single call. For a team building on an unproven voice front end, that is the hedge that matters: the speech layer can be swapped or stranded without taking the reasoning layer with it, and the model you bet on in March can be replaced in June by editing one line.

What to watch

Three things would change this comparison, and none of them are in anyone's hands yet. A Sesame API — even a paid one — would instantly turn the strongest voice in the category into a buildable option, and would put a real price next to GPT-Live-1's $0.05. A published benchmark of the Preview voice would convert a listening impression into a number, and independent evaluations tend to be unkind to voices that only ever competed in demos. And the first outside reproduction of Ope​nAI's interactivity and latency figures would settle whether full duplex is a real capability gap or a well-chosen evaluation.

Until then the honest split is this. If you are choosing a voice for yourself, use Sesame Preview — it is free, it is better, and it costs you nothing to prefer it. If you are choosing a voice for a product you have to ship, sell and support, GPT-Live-1 is the only one of the two you can actually buy, and the fact that it is not the better-sounding one is the price of admission.