AssemblyAI-TTS-1 vs Sesame Preview: One of These Has No Rate Card

You cannot buy Sesame Preview. It is a free consumer app — Maya, Miles, Simone and Charlie, four persistent-memory companions on iOS and Android, a 30-minute call cap and an unannounced plan to sell you glasses in 2027 — and there is no developer endpoint, no model ID and no price list behind it. AssemblyAI-TTS-1, by contrast, is a line item: $15 per million characters, generally available since 27 October 2026, streaming through HTTP and WebSocket in US-East or EU-North. So the matchup as usually stated is not a matchup. It is a product against a demo.

The comparison becomes useful the moment you substitute the Sesame that does have weights: Sesame CSM 1B, the Apache-2.0 conversational speech checkpoint on Hugging Face, a roughly one-billion-parameter model that accepts text plus audio context and emits English speech. Now it is a real decision — a 1B open checkpoint at $7 per million characters against a closed endpoint at $15 — and the interesting part is that the cheaper one is the one you have to build an agent around.

What Sesame Preview is, precisely, so the comparison does not slip

Sesame Preview is a shipped product with a real model inside it and no way to rent that model. The app's production model is not the released checkpoint: Sesame publishes CSM 1B, a base generation model that has not been fine-tuned to a specific voice, while the polished personas in the app are Sesame's own. The app claims sub-300 ms latency; third-party measurement of the Maya persona puts first-byte around 80–130 ms. Neither number describes CSM 1B as you will receive it.

The October 2026 update to the app — improved expressivity and pronunciation, agent computer access, Gmail, Calendar and Drive connectors, MCP support, an image generator — is a roadmap signal, not a commercial one. Sesame has said developer tooling is coming "later this year" without specifying what it will be, at what price, or whether it will be the production model or the checkpoint. Nothing in this decision should be built on that sentence.

The two bills, converted to the same unit

CSM 1B is listed at $7.00 per million characters on hosted gateways that serve it, billed per character rather than per token. AssemblyAI-TTS-1 is $15.00 per million characters, and that rate includes every voice in the 4,000-plus library, every one of its six launch languages, instant cloning, word timestamps, US or EU processing and unlimited concurrency.

At roughly 900 characters per minute of ordinary speech, the hourly comparison is about $0.35 for CSM 1B against about $0.81 for AssemblyAI-TTS-1 — the open checkpoint is a bit over twice as cheap per character — and the CSM figure is the one that tells you least, because it prices the model and not the system.

The dimension list is where the two stop being comparable at all:

• Price — $7 per million characters against $15, with all voices and cloning included on the higher one.

• Languages — CSM 1B is English, and the model card is blunt that any non-English ability is incidental; the app lists several languages as "early support", which is a different promise. AssemblyAI-TTS-1 launched with six.

• Voice inventory — CSM 1B ships seven voices and a base model you are expected to clone from; AssemblyAI-TTS-1 ships 4,000-plus and instant cloning from a 24-second sample.

• Cloning depth — neither offers a trained professional clone. AssemblyAI's instant path requires a speaker-permission confirmation step; CSM 1B transfers identity in zero-shot and flattens prosody doing it.

• What you must supply — CSM 1B needs a codec, turn detection, context assembly and the agent loop, plus a 4K-token context budget you manage yourself. AssemblyAI-TTS-1 needs an HTTP client, a stream handler and a cancel message.

• Latency evidence — AssemblyAI-TTS-1 has a vendor claim, two independent measurements and a board rank. CSM 1B has deployment reports that span two orders of magnitude.

Latency: 90 ms claimed, 159 ms measured, 9.9 seconds observed

AssemblyAI-TTS-1's latency story is documented and contested in a way that is easy to state honestly. The vendor claims 90 ms to first audio in production and 47 ms median time to first byte in its lab. Cekura, measuring independently in an agent-shaped test, got 159 ms median streaming startup, 296 ms at p90, and ranked it eighth of the nineteen models it timed. That gap is methodology — request-to-first-byte versus last-word-sent-to-first-audio — not dishonesty, and the second measurement is the one your caller experiences.

CSM 1B's latency story is not contested so much as unbounded. Optimised on-device conversions report about 120 ms to first audio with a real-time factor between 0.5 and 0.9 on Apple Silicon; an unoptimised RTX 5070 Ti reportedly runs at roughly half real time; one hosted listing publishes a P50 of about 9.9 seconds and a P95 of about 39 seconds end-to-end. All of those describe the same checkpoint. Which one you get is a function of how much serving engineering you bring.

That is the crux of the whole comparison, and it is not a quality argument. A closed endpoint has already done the serving work and priced it in: $0.81 an hour, no GPU, no codec, no queue design. An open 1B checkpoint gives you a model for $0.35 an hour and hands you back the part that was expensive to build. If you already own that engineering, the checkpoint is the better deal by a wide margin. If you do not, the $0.46 an hour you save is a rounding error next to the first month of someone's time.

Where each one is actually bad

AssemblyAI-TTS-1's weak spots are all in the tail of real text. Artificial Analysis scores it 81.7% on pronunciation robustness overall, sixteenth, with shorthand expansion at 73.8%; Cekura's contextual pronunciation test puts it twenty-second of twenty-seven at 70.8%, even while ranking it first on phone numbers and codes at 91.7%. There is no pronunciation dictionary and no SSML to fix any of that, and its six languages are a hard ceiling rather than a starting point. Vendor-reported entity accuracy of 93.7% on its own 580-item set says the opposite, and only your corpus can settle which measurement describes your traffic.

CSM 1B's weak spots are structural rather than data-driven. It is English-only by design. It is a base model, so the voice you get is not a voice anybody tuned. It clips the final word on some generations, carries a watermark in its reference generator, and some hosted wrappers silently drop the context, language and max-length parameters — which means a gateway quoting you CSM 1B may not be running the model as documented. Its context is 4K tokens with a maximum output around 3,686, so long conversations have to be assembled by you.

One thing neither of them is, and this matters for anyone coming from the consumer app: a conversational partner. AssemblyAI-TTS-1 reads text into speech and has no notion of dialogue state. CSM 1B is the more dialogue-aware of the two — it takes audio context as input — but it produces speech one turn at a time with no turn detection of its own. The app's apparent seamlessness is a product, and neither model ships it.

Three questions people actually ask

Is CSM 1B the same model the Sesame app runs? No. The app runs Sesame's production model; the checkpoint is a base generation model released under Apache 2.0, and the gap between them is exactly why the app sounds like Maya and a fresh CSM 1B deployment does not. Treat the app as a demo of what Sesame can build, not as a preview of what you will get.

Which is cheaper at a hundred thousand three-minute calls? On paper, CSM 1B: 270 million characters, about $1,890 on the checkpoint against about $4,050 on AssemblyAI-TTS-1. In practice the checkpoint figure excludes GPU hours, the serving engineer, and the latency work needed to reach anything close to an agent-grade stream. At that volume the honest comparison is $4,050 against "unknown until you have built it", and unknown is the more expensive of the two.

Can AssemblyAI-TTS-1 clone my own voice? Yes, by uploading a recording of around 24 seconds and confirming you have the speaker's permission; the voice ID comes back within seconds and bills at the same per-character rate as the library. There is no trained professional clone, so the result is a usable identity rather than a governed brand asset — and no fine-tuned cloning was available at launch.

Which decision this actually is

If the name in front of you is Sesame Preview, stop: there is nothing to integrate and no price to compare. If it is CSM 1B, then you are not choosing a TTS model, you are choosing whether to own the serving layer — and the three conditions that make "yes" defensible are that you need English only, that you have GPU capacity with an engineer to match, and that you want the model inside your own perimeter. Everything else argues for the endpoint.

The cost argument that usually decides it is not the per-character rate at all. It is consolidation: a voice agent tends to accumulate a TTS vendor, a speech-to-text vendor and a language-model vendor, each with its own contract and its own price-change cycle. OrcaRouter exists to flatten that last layer — one API for 200-plus models at provider list price with 0% markup, so a vendor's price cut reaches your bill the same day rather than at the next renewal, with automatic failover, a routing DSL for composing models into one call and model fusion when a single answer will not do. AssemblyAI-TTS-1 and CSM 1B are both outside that catalogue today — our speech routes are OpenAI's tts-1 family and Google's Gemini TTS previews — so the argument is about the bill you are already paying, not about a purchase you can make here.

Sesame's app is the best free demonstration in this category and will probably get a developer product eventually. AssemblyAI-TTS-1 is a rate card you can sign this week. Between those two facts, the only defensible plan is to build on the one that exists, and leave a slot in your architecture for the other if Sesame ever ships the model instead of the brochure.