Rime Prism vs Sesame Preview: A Rate Card Against a Reputation

You can hear Sesame Preview tonight for nothing. You cannot buy it. You can buy Rime Prism at $0.080 a minute, and almost nobody outside Rime has published a clip they were allowed to measure. Both halves of that sentence are the comparison: one of these is the best-regarded voice in consumer AI that no developer can call, and the other is a brand-new enterprise endpoint that a startup can put in a phone tree this week. Rime Prism launched on 28 October 2026 as Rime's first speech-to-speech model. Sesame Preview has been running as a free research preview since early 2025, with the production voice behind it still unreleased.

They get compared because both make a claim about conversational realism, and because both are English-only at launch. They are not alternatives. One is an application you talk to, the other is an endpoint you build on, and the decision between them is less a choice of voice than a choice of whether you are buying a product or a component.

The split that decides almost everything

A team that wants to ship an agent asks two questions in this order: can I call it, and what does it cost per call. Sesame Preview fails the first. There is no public API for the model behind the app, no model ID, no endpoint and no published price. The Sesame you can talk to is a consumer product with four named assistants, persistent memory and a growing set of tools; the model underneath it is not something a developer can address.

The only Sesame thing anyone can call is CSM-1B, the open-weight checkpoint the company released in 2025 under Apache 2.0: roughly 1B parameters, a 4K context, 24 kHz mono audio. That is not the voice in the app, and it is not an agent — it conditions on conversation history but cannot reason about it the way a language model does. It is a speech component, and a single one at that.

• Availability — Rime Prism is a priced API with a free tier; Sesame Preview is a free consumer app with no API for its production voice.

• Price — Prism $0.080 per call-audio minute on Flex, $0.064 on Scale; Sesame publishes no price and charges nothing for the app.

• Billing risk — Prism bills whether the call goes well or not; Sesame's cost is whatever the app costs, which is zero and not a business model you can plan around.

• Latency basis — Prism 0.57 s time-to-first-audio, vendor-measured on τ-Voice; Sesame's app claims sub-300 ms, and third-party measurement of the Maya persona has put time-to-first-byte at 80 to 130 ms.

• Independent score — neither model appears on Artificial Analysis's Speech-to-Speech Index.

• Languages — both English-only at launch; Sesame lists French, Italian, German, Spanish, Chinese, Japanese and Korean as early support.

• Deployment — Prism offers cloud, VPC and on-premises with a HIPAA BAA and SOC 2 Type II; Sesame offers an app.

Rime Prism, as an endpoint

Prism is sold the way enterprise software is sold: a free tier with $50 of monthly credit and no card, Flex at $0.080/min with no commitment, Scale at $100 a month with $100 of credit and $0.064/min, and Enterprise with volume rates and annual commitment. Every tier is the same model. Prism streams over WebSockets and plugs into LiveKit and Pipecat, which matters because those two are where most self-built voice agents already live.

The features that justify the rate are the ones a consumer app has no reason to build. Prism writes its reply as text before speaking, so an application can read the exact sentence, run a policy check or a PII filter against it, and let the audio proceed only if it passes — while the model keeps listening, so the check does not have to become an audible pause. Tool calls run in parallel with speech. Fine-tuning on a customer's own calls and tool traces is the core business case, with self-serve SFT and RL APIs still in beta.

The company's own numbers put time-to-first-audio at 0.57 seconds and τ-Voice completion at 56.6%, with 82.5% on the benchmark's telecom subset. A Revmo case study on around ten hours of real ordering calls is cited at 100% order confirmations against 68% for an unnamed OpenAI model. That is a commissioned customer study with no published methodology, and it should be read as a vendor claim rather than a result.

Sesame Preview, as a product

Sesame's pitch is that a voice assistant should feel like a person you know. The four personas — Maya the witty collaborator, Miles the relaxed one, Simone for deep dives, Charlie as a casual sounding board — each carry their own memory and their own way of talking. The app searches the web, runs long research, keeps notes and reminders, summarises calls and supports an incognito mode. On 6 October 2026 Sesame shipped an updated CSM with better expressivity and pronunciation, gave its agents access to their own computer, and added Gmail, Calendar and Drive connectors, external MCP servers over OAuth, user-built skills and an in-app image generator.

What it does not have is a way in. The app is free, runs no ads and does not sell data, which is a pleasant thing for a user and an unresolved thing for a buyer. How Sesame intends to make money, and whether developers ever get an endpoint, is still unannounced. The glasses are dated to 2027 and Japan; developer tooling has been promised for "later this year" without specifics.

The reputation is real, though. Sesame's voices are the ones people send each other clips of, and reviewer consensus has been unusually warm for two years running. The problem is that a reputation is not a specification: no latency figure has been published for the production model, no blind-listening score exists for it, and the only independently runnable piece is the 1B checkpoint, which is a different thing.

The one capability Sesame has ruled out

If your reason for wanting Sesame's voice is that it sounds like a person, the thing you actually need to replicate is a conversational loop, not a timbre. And a conversational loop for a business has requirements a consumer app never has to meet: an audit trail of what the agent was about to say, a way to stop it saying it, a tool call that resolves before the caller notices, and a deployment that clears a compliance review.

Prism is built around those requirements, and that is why it costs money. Sesame is built around the opposite: a closed, curated, free experience in which the company controls every variable. Both are defensible products. Only one of them is a component you can wire into a claims line.

Both are English-only — for opposite reasons

Prism launched English-only with a limited voice set because it is a first release and voice coverage is expensive to ship. Sesame is English-first because its product is a set of English-speaking personalities; French, Italian, German, Spanish, Chinese, Japanese and Korean are labelled early support, which is a different promise from "supported."

For a multilingual contact centre this is the section that ends the article. Neither model covers a mixed-language queue today, and the comparison that actually matters for that buyer is against a TTS layer with broad language coverage plus a routed text model, not against either of these.

The measurement problem

Neither model is on Artificial Analysis's Speech-to-Speech Index, the one independent board that ranks conversational voice models on task completion, arena preference and task success. That absence is not a knock on either product; it is the reason this page has to lean on vendor claims for both.

Prism's 56.6% τ-Voice figure is Rime's own run. Read the table Rime publishes next to it and GPT-Live-1 (Sol, low) appears at 59.3% — higher than Prism — while the copy claims Prism is ahead of every OpenAI and Google realtime model. Launch posts also repeat "under 300 ms" and "4× faster," numbers that do not appear in the benchmark table, which says 0.57 s. Those may be measurements of different things; Rime does not say which.

Sesame's position is harder still, because there is no production model to measure. It has, to its credit, built the tool that would measure one: TurnBench, released on 28 August 2026, is an open-source benchmark for turn-taking in spoken dialogue, built with Carnegie Mellon, Mundo AI and Oto, with a public leaderboard and self-serve scoring. It is Sesame's own benchmark rather than a neutral one, but it is a genuine contribution, and the number it is built around is the most useful figure in this whole comparison: the typical gap between human turns is about 200 milliseconds, while planning a single word takes about 600. That is why turn-taking, not voice quality, is what gives a machine away.

The only independent evidence attached to Rime's catalogue is its TTS line, and on the Text-to-Speech Arena Coda ranks 65th of 99 at Elo 1,054 and Mist V2 92nd at 903. That says nothing definitive about Prism, whose voices are new — but it does mean "Rime has never had a top-ranked voice on an independent board" is a true statement a buyer should hear before signing a Scale contract.

What a phone line costs each way

Prism is easy to price. Ten thousand minutes of inbound calls is $800 on Flex or $640 on Scale, before any Enterprise discount. A five-minute call is $0.40. If the agent also needs to summarise the call afterwards, that is a text-model cost on top and it is small next to the voice bill.

Sesame is impossible to price, because there is no rate to price. The app costs nothing and cannot be embedded; the open checkpoint costs $7 per million characters from its only hosted provider, and self-hosting it means owning the inference stack, the codec, the turn detection and the agent logic yourself. Anyone quoting a per-minute cost for "Sesame" is quoting the app's absence of a bill, not a service.

That is the practical asymmetry: Prism is expensive and legible, Sesame is free and unavailable. A budget meeting can argue with $0.080 a minute. It cannot do anything with "no endpoint."

Three buyers, three different answers

An enterprise running inbound support calls should look at Prism. The text checkpoint, the HIPAA BAA, the on-premises option and the per-minute rate are all pointed at exactly that job, and the vendor-claim risk is survivable because the commercial terms are ordinary. Pilot it on one call flow, measure the containment rate yourself, and treat Rime's 100%-versus-68% case study as a hypothesis to test rather than a forecast.

A developer who wants Sesame's sound inside a product should not wait for an API that has not been announced. The realistic path is an open TTS checkpoint plus your own orchestration — and the orchestration is the part you can rent cheaply. OrcaRouter puts 200-plus text models behind one key at provider list price with 0% markup, adds automatic failover so a voice agent does not die when one provider has a bad minute, and exposes a routing DSL so routine turns can go to a small model and escalations to a large one. The voice layer and the reasoning layer do not have to come from the same company, and for most builders they should not.

A consumer, or anyone building a personal assistant for themselves, should just use Sesame. It is free, it is good, and the absence of an API is irrelevant to someone who is not writing code.

What would change the answer

Three events. Sesame publishing an API and a rate card would turn this into a real comparison overnight, and would immediately raise the question of whether its voice quality survives being sold as a component. Prism appearing on the Speech-to-Speech Index would give it the first independent number it has ever had, and would settle whether the "ahead of OpenAI and Google" line survives contact with a neutral board. And Prism's fine-tuning APIs leaving beta would separate the enterprises that can already buy it from everyone else.

Until one of those happens, the comparison is between a product you can experience and a product you can deploy, and the only wrong answer is treating them as two versions of the same purchase. Sesame is the voice that made people care about conversational AI again. Prism is the first serious attempt to sell that feeling by the minute with an audit log attached. Neither of them is the other one's replacement.