Grok Voice 3.0 vs Sesame Preview: The Voice That Wins Every Listening Test Is the One With No API

Ask a room of people which of these sounds more like a person and you will get one answer. Sesame Preview is the free voice assistant from Sesame — four agents named Maya, Miles, Simone and Charlie, on iOS since May 2026 and broadly on Android since early August — and it produced the viral clips that made "best-in-class voice mode" a phrase reviewers keep reaching for. G​rok Voice 3.0, announced by S​paceXAI on 14 October 2026, is the highest-scoring speech-to-speech model on the public board at 86.9, callable over a documented realtime endpoint at $0.08 a minute. Those two facts do not contradict each other, and the gap between them is the whole matchup: one of these is a voice people judge with their ears, and the other is a voice an outside party can measure.

The second half of that gap is access. Sesame Preview has no API, no model identifier, no rate card and no stated plan for one, and the production voice in the app is a larger closed model the lab has never released. G​rok Voice 3.0 is a component you rent by the minute. Pick by ear and you want Sesame. Pick by whether you can ship something and the decision was made for you.

What each one is, on the record

Stripped of the marketing, the two products are not even the same kind of thing.

• What it is — G​rok Voice 3.0 is a speech-to-speech model behind S​paceXAI's realtime WebSocket, with session-scoped tools and returned transcripts; Sesame Preview is a consumer app with four conversational agents, per-agent memory that syncs across web and mobile, live web search during a call, and a summary sent afterwards.

• Price — $0.08 per minute of audio plus $0.004 per text input, about $4.80 an hour, for G​rok Voice 3.0; free to the user with no ads and no stated paid tier for Sesame Preview.

• Developer access — a documented endpoint, ephemeral tokens for client-side connections, and a custom-voice endpoint that trains a voice from a reference clip of up to 120 seconds for G​rok Voice 3.0; nothing at all for Sesame Preview.

• Weights — closed on both sides, with one exception: Sesame has published CSM-1B, an Apache-2.0 checkpoint that is not the voice in the app.

• Independent score — 86.9 on the Artificial Analysis Speech to Speech Index with five readable components for G​rok Voice 3.0; none for the Sesame Preview voice, and none published by Sesame either.

• Limits — G​rok Voice 3.0 has no stated session cap, and its predecessor generation was marketed on word error rate across 24 languages; Sesame Preview is English only, with a 30-minute call cap that drops to five minutes when logged out, and no task execution — the agents talk, they do not book, buy or file anything.

That last bullet is the one that ends most evaluations. A voice product that cannot be called is not a component; it is a competitor.

The open checkpoint, and what it is not

Sesame has been unusually direct about the gap between its demo and its download, and it is the single most important fact for anyone weighing the two.

CSM-1B is a 1B conversational speech model released under Apache 2.0 on 23 April 2026: a L​lama-architecture backbone feeding a separate decoder and Kyutai's Mimi codec, producing 24 kHz mono English speech from a context of roughly 4,000 tokens. It is a base speech generation model. By the assessment of people who have run both, it is a distinctly weaker voice than the one in the app, and Sesame's own documentation is explicit that it is not an end-to-end speech-to-speech system and cannot generate text.

So the download is a research lineage, not the product. Someone who went looking for the voice from the demo found the smaller model that came out of the same lab, and the thing that answers when one of the four agents picks up is a production model that has never been released. It is hosted by third-party inference providers at around $7 per million characters — roughly $0.35 an hour of synthesised audio — or free on your own hardware, and nobody has published an independent benchmark of it either.

For scale, that $0.35 an hour is the cheapest audio in this comparison by a factor of thirteen, and it is also the least useful line in it, because a 1B checkpoint that cannot hold a duplex conversation is not competing with G​rok Voice 3.0 on price. It is competing with whatever narration tool you were already using.

Why the ear test and the index disagree, and which one is wrong

They are not in conflict. They are answering different questions, and only one of them can be answered in public.

Sesame Preview's reputation rests on how a voice sounds to a listener: breath, hesitation, the timing of an interruption, the impression that a person is on the line. That is real evidence of a kind — the demo persuaded hundreds of thousands of people and reviewers who had no incentive to be generous — and it is completely unquantified. There is no arena entry, no published evaluation of the production voice, and no sample count. Every quality claim in either direction is a listening impression.

G​rok Voice 3.0's 86.9 is the opposite kind of evidence. It decomposes into speech reasoning near the ceiling, 97.0% on Full Duplex Bench, 71.6% on τ-Voice, an arena Elo of 1070 and 94.8% task success, each measured under a published harness that an outside team can read, check the arithmetic on and disagree with. A composite index is a weighting choice, and the arena component is frozen at the point a model becomes eligible rather than tracking live Elo — but the disagreement is available to anyone, which is what makes it a measurement rather than a claim.

What the board does not measure is whether you enjoy talking to it. And there is a specific number in the release that lands right on that axis: G​rok Voice 3.0 starts speaking in 0.81 seconds, and the model it retires started in 0.70. The flagship is 0.11 seconds slower off the mark than its own predecessor, which is precisely the kind of detail that a listener notices before a benchmark does.

What the four-agent design tells you

Sesame's four personalities are not a marketing detail; they are an architectural decision, and it is the opposite of the one S​paceXAI made.

Maya, Miles, Simone and Charlie are fixed voices with fixed temperaments, and memory is scoped per agent. That is a product built around the assumption that the voice is the deliverable and the personality is the feature. There is no voice customisation, no cloning, and no path from the consumer app to an API — the free tier is the product, and the company's stated direction is research partnerships and a planned hardware device rather than a developer platform.

G​rok Voice 3.0 goes the other way. It ships with S​paceXAI's built-in voice set, and a separate custom-voice endpoint trains a voice from a reference clip of up to 120 seconds, with the resulting voice_id working across both the realtime API and streaming text-to-speech. Its session accepts tools, so the agent can call your backend mid-conversation rather than only talking about it.

Which one is better depends on whether you are selling an experience or building one. A voice product that ships four excellent fixed personalities has solved branding and integration in one decision. A voice product with a cloning endpoint has handed you a compliance surface and a localisation workflow in one decision. Neither is a mistake, and they point at different customers.

Where the routing question actually lands

OrcaRouter serves neither of these voice layers. G​rok Voice 3.0 lives behind S​paceXAI's own realtime WebSocket and nothing can interpose on a duplex audio session owned by one vendor; Sesame Preview has no endpoint to route to at all. Saying otherwise would be describing a capability that does not exist, and this category is full of platforms that do exactly that.

What is routable is the layer both products quietly depend on. Sesame's agents search the web mid-conversation and send a summary afterwards; G​rok Voice 3.0's tools array reaches whatever backend answers it, and anything heavier than an inline call becomes a text-model request behind the conversation. That work is request-metered, it is where a meaningful slice of per-call cost accumulates, and it is the half of a voice product you can hold constant while the audio model underneath it changes — which, given that grok-voice-latest/ has now moved twice in one quarter, is a property worth having. Two hundred and seven models from a dozen-plus upstreams behind one key, at provider list price with zero markup, with automatic failover when a provider degrades, a routing DSL for composing several models into one call, and model fusion where one model's judgment is not enough. Because nothing is marked up, a vendor price cut on any of those models is live here the same day.

The recommendation, split by who you are

If you are a consumer who wants the best-sounding conversational voice available today and has no intention of integrating anything, Sesame Preview is free, it is on both mobile platforms, its reputation is deserved enough that reviewers keep confirming it, and nothing in this article changes that. Use it and enjoy it.

If you are building, the decision is already made and the reasoning is worth internalising. You cannot call Sesame Preview, you cannot pin its voice, and you cannot benchmark it against anything — so every procurement conversation about it ends in the same place, which is a listening test between two options neither party can score. G​rok Voice 3.0 gives you a rate card, a versioned model string, a transcript, a cloning endpoint and a board score you can argue with. The first thing to do with it is pin your version, because the alias has already been moved onto it without asking.

And if you are a researcher who wants the interesting failure: the honest summary of this pairing is that the category has proved it can make a voice worth listening to and has not yet decided whether that voice should be purchasable in any form other than an app. Sesame built the better experience and kept it closed. S​paceXAI built the better component and published the number. Watch for the version of Sesame that has an API — that is the release that would make this page obsolete.

That gap between what a scoreboard measures and what a deployment meets is the reason the record matters: the documented AI failures and agent incidents around models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.