Fish Audio Live 3 vs Sesame Preview: Two Answers to Turn-Taking, One of Them Measurable

Sesame built a benchmark for the exact problem Fish Audio Live 3 says it solved, and then Fish Audio shipped a model that is not on it. That is the short version of this comparison. Sesame Preview — the four-agent consumer product running on the lab's Conversational Speech Model — published TurnBench in August 2026: thirty hours of triple-annotated two-person speech, a public leaderboard, an interactive conversation viewer, and results for fourteen systems. Fish Audio Live 3, announced on 10 November 2026, argues that conversational timing should be learned inside the speech model rather than imposed by silence thresholds, and publishes three sample calls and a 186-millisecond figure. The argument is the same argument. The evidence is not the same evidence.

Everything else follows from how differently the two companies turned that argument into a product. Live 3 is sold: $0.035 per minute on Fish Audio's Agent Platform during a promotion that ends 31 December 2026, a documented API, and an on-premises option for enterprises. Sesame Preview is given away: four fixed personalities in an iOS and Android app, with no API, no model identifier for the voice in the app, and no per-token or per-minute anything. One of them you can build on this week. The other one you can only listen to.

The test that exists, and why Live 3 is not on it

Turn-taking is the rare capability where the industry agreed on a way to measure it before the models arrived, and Sesame is a large part of why. TurnBench, released on 28 August 2026 and built with Carnegie Mellon University, Mundo AI and Oto, scores systems on how they handle exactly the events that make a phone agent feel wrong: an acknowledgement that should not steal the floor, an interruption that should, a pause that is not an ending. The leaderboard is public, the development set is self-serve, and the results cover fourteen systems.

Fish Audio Live 3 is not among them, and neither is any other Fish Audio duplex model. That is not a scandal — a model released in November cannot be on an August leaderboard — but it is the single most useful fact in this comparison, because it tells you what kind of claim each company is making. Sesame put its own product on a public test it helped design, where anyone can check the score. Fish Audio put its product on a promotional rate card, where anyone can check the price.

Both are legitimate launches. Only one of them lets a stranger verify the thing the marketing is about.

What each side actually ships

• What it is — Fish Audio Live 3 is a full-duplex speech model behind an agent platform; Sesame Preview is a consumer assistant app with four agents, Maya, Miles, Simone and Charlie, each with its own memory.

• Distribution — Live 3: self-serve platform, documented API, enterprise on-premises. Sesame Preview: iOS and Android apps, free, no API, no model ID for the voice in the app.

• Price — Live 3: $0.035/min promotional, $0.07 list, plus $0.01/min for phone calls and $1.20/month per number. Sesame Preview: no published price and no published plan tier; the app is free to download.

• Languages — Live 3: English only at launch, in Fish Audio's own words, even though the surrounding platform advertises voices in 30+ languages. Sesame Preview: English-centric, trained mainly on English according to the lab's own research notes.

• Independence — Live 3: no published benchmark of any kind as of 11 November 2026, and no row on the Artificial Analysis speech-to-speech board. Sesame Preview: no independent score for the app's voice either, but a public turn-taking leaderboard the company helped build and publishes.

• Openness — Live 3: no weights, no licence, no repository. Sesame: one open checkpoint, CSM-1B, Apache-2.0 on Hugging Face since March 2025 — which is not the voice in the app.

• Interface — Live 3: you integrate it. Sesame Preview: you talk to it.

The vocabulary problem both companies share

Neither product publishes enough to let you compare the thing you actually care about, and both use the same word for it. Fish Audio calls it the rhythm of conversation and describes a phone number read in chunks — "415… 555… 0132" — where a conventional agent hears three separate ends of turn. Sesame's marketing calls the same behaviour presence, and its October 2026 release notes say the company updated its Conversational Speech Model to improve expressivity and pronunciation, with less sycophancy and fewer repeated phrases. Both statements are unfalsifiable as written. Neither ships a number.

The difference is that Sesame shipped the ruler. If Fish Audio wants the comparison on the merits, TurnBench's development set is open and self-scoring; the scoreboard is waiting for an entry.

What a builder can do with each, and what they cannot

This is where the two products stop being comparable at all. Sesame Preview is not a component. There is no endpoint, no session API, no way to put its voice behind your own phone number, and the only Sesame checkpoint anyone outside the app can load is CSM-1B, a 1B-parameter model from March 2025 that predates the voice in the current app by a year and a half of work. A team that wants Sesame's conversational feel in a product has, today, no path to it that does not involve waiting.

Live 3 is the opposite trade. It is closed in the way that matters — no weights, no way to host it yourself outside an enterprise contract — but it is open in the way that lets you ship: an agent builder, a documented API, telephony, and a per-minute rate. The reasoning layer is a separate, replaceable part; the agent can point at an OpenAI-compatible endpoint for the thinking while the speech stays Fish Audio's. That structure is what makes it deployable by a small team on a Tuesday.

For the reasoning half specifically, OrcaRouter carries more than 200 models behind one key at provider list price with zero markup, with automatic failover when a backend errors mid-call and a routing DSL for deciding which model handles which kind of turn. Fish Audio Live 3 is not one of those models — there is no third-party endpoint for a closed duplex speech model, and there is nothing here to route. The router's job in this stack starts where the speech ends.

Which one to watch, and for what

Three signals would settle this comparison, and none of them is a press release. A Live 3 entry on TurnBench or a row on the Artificial Analysis speech-to-speech board would turn Fish Audio's behavioural argument into a score with a number next to it. A Sesame developer API — the company has said developer tooling is coming, without a date — would turn the app's voice into something a product could rent. And a second open Sesame checkpoint, newer than CSM-1B, would give the open-weights crowd the thing that Live 3's launch conspicuously did not.

Until one of those lands, the honest ranking depends entirely on what you are buying. If you want the best-sounding conversation available today and you are willing to have it inside someone else's app, Sesame Preview is the more impressive product and the more honest one about how it should be judged. If you need a voice in production this quarter, Sesame is not a candidate and the comparison collapses into a single question about Live 3: whether a $0.035 promotional minute and a 186-millisecond acknowledgement floor are enough evidence to bet a phone line on. For most teams the answer is yes for a pilot and not yet for a payroll — and the benchmark that would change that answer already exists, with a row left blank.