StepAudio 3 Realtime vs Sesame Preview: The Voice Everyone Praises vs the One That Has a Score

For most of 2026, the strongest claim anyone could make about voice AI was a listening impression. Sesame's assistant sounded more human than anything else, and enough people said so that it became the reference point — with no benchmark, no number, and no way to check. On September 15, 2026, StepAudio 3 Realtime arrived from StepFun with a number attached to the closest measurable version of that quality: 98.9% on Artificial Analysis's Conversational Dynamics evaluation, first place. Sesame Preview is not on that board.

The interesting part is not that one beat the other. It is that the quality Sesame became famous for is now something a third party scores, and the model with the reputation has never been measured against it.

What each of these actually is

They are not the same kind of object, and that determines more of the comparison than any score.

Sesame Preview is a consumer voice assistant. Free on iOS since May 2026 and broadly on Android since early August 2026, built around four named agents — Maya, Miles, Simone and Charlie — that hold context across sessions, search the web and set reminders. It is English-only. Calls run to a 30-minute cap, which drops to five minutes when you are logged out. There is no API for its production voice, no enterprise tier and no public pricing.

StepAudio 3 Realtime is a developer surface. It is one of five models in the StepAudio 3 family, all released the same day and all live on StepFun's open platform as commercial APIs. It handles native full-duplex conversation, reasons directly over speech rather than transcribing first, and supports tool calling and asynchronous execution of long tasks while the conversation continues. It is Chinese-first. It is not open weights, and it is currently on limited-time free preview pricing with no post-preview rate announced.

One of these you can build on. The other you can visit.

The measurable thing Sesame is famous for

Conversational Dynamics is a specific benchmark, and it is worth knowing what it holds. It scores turn-taking, pause handling, interruption recovery, and whether a model correctly reads a brief backchannel — "uh-huh", "right", "mm" — as encouragement rather than as a bid for the conversational floor. Those are precisely the behaviours listeners describe when they say a voice feels human rather than robotic, and they are the behaviours that make an assistant pleasant to talk to rather than merely accurate.

StepAudio 3 Realtime leads that board at 98.9%, ahead of Qwen Audio 3.0 Realtime Plus at 98.4% and GPT-Realtime-2 at 95.3%. It also places first on Speech Reasoning at 99.7% as cited by StepFun, behind which lies the architectural claim that reasoning over raw audio preserves tone and emphasis a transcription hop would flatten — the difference between hearing sarcasm and hearing a compliment.

Here is the honest framing of that result. The benchmark is the closest thing the industry has to a measurement of what Sesame is praised for, and Sesame has not been entered into it. That is not evidence that StepAudio 3 Realtime sounds better than Sesame. It is evidence that one of these claims is checkable and the other, so far, is not — and that the checkable one is currently held by a model most people have never spoken to.

The availability asymmetry, which is the real decision

Everything above is interesting. This part is decisive.

Sesame's voice — the one people rave about — is a larger, closed production model that Sesame has never released as an API. You cannot buy it, license it, or call it from your own product. If you read that Sesame's conversational timing is the best available and concluded you could have it, the conclusion does not follow from the evidence.

What Sesame has actually open-sourced is CSM-1B, released under Apache-2.0. It is a synthesis block, not an assistant: a Llama backbone predicts the first Mimi codebook and a smaller Llama decoder predicts the rest, which a Mimi codec renders to a 24 kHz waveform. It is English-only, it clones a voice from a 10-to-15-second reference clip, and it cannot generate text at all — you pair it with a separate language model. People who have run both describe CSM-1B's voice as distinctly weaker than the Preview app's, and the model card says the quiet part out loud: a fine-tuned variant of CSM powers the interactive demo, and that variant is not the checkpoint you can download.

StepAudio 3 Realtime has none of that ambiguity. It is a commercial API with a published model family behind it, a stated capability set, and third-party placements you can look up. It is also Chinese-first, preview-priced, and posts a time to first audio of 8.83 seconds on the same leaderboard where it wins conversational dynamics — against 1.18 seconds for Gemini 3.8 Live and 1.24 to 1.34 seconds for GPT-Live-1. Whatever it offers in conversational quality, it has not yet demonstrated that it can deliver it at conversational speed outside StepFun's own serving setup.

What to do with this if you are choosing a voice stack

Start by separating the two questions. "What should conversation feel like?" and "what can I ship?" have different answers, and only one of them is a procurement decision.

If you are calibrating taste — deciding how much warmth your product should have, how much silence is comfortable, how quickly an agent should yield the floor — Sesame Preview is free, genuinely excellent, and a good place to spend an afternoon. Use it as a reference point. Do not plan an integration around it, because there is nothing to integrate with.

If you are shipping, StepAudio 3 Realtime is a real candidate with real caveats: preview pricing, Chinese-first coverage, no index score on Artificial Analysis, and the latency question unresolved. Test the latency before the conversation quality. A model that wins the turn-taking benchmark but takes most of a sentence's worth of time to start speaking is a model whose best quality you may never get to demonstrate.

And if Sesame's production voice ever gets an API, this whole comparison gets re-run — because the benchmark that would settle it already exists, and Sesame would finally have to appear on it.

Where routing fits, and where it does not

Voice agents are not one model. Whether you are running StepAudio 3 Realtime or anything else, the conversation constantly hands work to ordinary text models — a lookup, an extraction, a reasoning call running behind a held pause. That delegation layer is where cost variance lives and where one provider's bad hour becomes your outage.

To be precise about our own coverage: the live and voice endpoints in the StepAudio 3 family are not on OrcaRouter, and neither is anything Sesame has built. What is on OrcaRouter is the text layer a voice agent delegates to — nearly 200 models behind one API, automatic failover, a routing DSL that composes several into one call, and model fusion when a single model's judgment is not enough, with provider list price passed through at no markup.

That split is what makes the availability question less fatal than it looks. The voice model is a decision you can revisit; the delegation layer is the one that gets expensive to unpick, and it is the one worth putting behind a router before you commit to any voice vendor.

The verdict

Sesame Preview wins on the thing that cannot be measured and loses on everything that can be bought. It is the best-sounding voice available, it is free, and you cannot put it in production — and now that a third party scores the quality it is famous for, its absence from that scoreboard is a gap in the evidence, not a protected lead.

StepAudio 3 Realtime wins on availability, measurability and capability, and carries genuine open questions on price, language coverage and latency. It is the only one of the two you can build on today, which settles the practical question faster than any benchmark.

What to watch is simple, and it cuts both ways: a Conversational Dynamics score for Sesame's production voice, or a latency figure for StepAudio 3 Realtime that puts it in the same range as the models it currently beats on quality. Either one would turn this from a comparison between a measurement and a reputation into an actual head-to-head.