Simba 3.5 vs Sesame Preview: One Is Sold by the Character, the Other by the Seat

Here is a comparison that looks like apples and oranges until you price it, at which point it becomes arithmetic. SpeechifyAI Simba 3.5 is a text-to-speech model you buy by the million characters and call from your own code. Sesame Preview is a consumer voice agent you buy by the month, per person, in an app. Both are purchasable — Sesame published a Free, Plus and Pro tier list and stopped being a free demo — and both produce a synthetic human voice. One is a component and one is a product, and the only honest way to choose between them is to work out which of those two things your problem actually needs.

Sesame the company does publish a model: CSM-1B, open weights on Hugging Face under Apache 2.0. It is not the voice you hear in the app. The app runs a production system with no API, no model id and no independent arena score, and the research release has not moved since 2025. That gap is the single most important fact in this piece and it is the reason the two products are not substitutes.

What each one costs, converted to the same unit

The unit mismatch is where most write-ups stop, so start there. Simba 3.5 bills per character of input text. Sesame bills per seat per month with a talk-time allowance. Neither rate converts cleanly into the other, but both convert into "what does this cost my organisation in a month," and that comparison is not close.

• Simba 3.5, list from 22 November — $15.00 per million characters on Starter, $12.00 on Pro, $10.00 on Scale

• Simba 3.5, through 21 November — $10.00, $8.00 and $6.00 per million characters, the older models' rates

• Sesame Free — $0, 30 minutes of talk time a day

• Sesame Plus — $20 a month, two hours of talk time a day, four times Free's weekly credit allowance

• Sesame Pro — $100 a month, ten hours of talk time a day, twenty times Free's weekly credits

• Sesame add-ons — $5 for one extra hour of talk time, $20 for five; credits at $5 per thousand and $20 per five thousand, packs that do not expire

Ten million characters of Simba 3.5 on the Scale plan is $100 and renders somewhere in the region of two hundred hours of finished speech. Sesame Pro is $100 a month and buys ten hours of talk time a day for one person. Those are different products with the same invoice line, and it is genuinely difficult to say one is better value, because one renders an asset and the other holds a conversation.

The Sesame number worth flagging is the overage rate rather than the subscription. Extra talk time at $5 an hour is roughly $0.083 a minute, which is expensive for a conversation and cheap for a product you have decided to put in front of customers. The packs never expire and are spent after the plan allowance, so the marginal hour is the number to model, not the headline tier.

What Simba 3.5 is, precisely

Simba 3.5 launched on 21 October 2026 as Speechify's recommended model, replacing Simba 3.2 as the go-forward English choice and extending coverage from six languages to 48 under a single model id. Every catalogue voice speaks all 48, including a cloned voice on a paid plan, with the same voice id. Estonian and Cantonese are explicitly excluded. Voice ids, output formats, speech marks and the applied SSML subset carry over from 3.2, so migrating is a one-line change to the model field with no API version bump.

On Artificial Analysis it is first on Controlled Voice at 1,192 and fifth on Provider Voice at 1,294, from 1,783 and 1,732 samples respectively. The controlled board is the meaningful one for this comparison: it clones the same eight reference voices for every entrant, so a model that scores well there handles a voice it has never heard. That is precisely what a cloned-voice workflow demands.

Speechify's own numbers, unreproduced by anyone outside the company: 63% blind preference over Simba 3.2 with a stated 95% interval of 60% to 67%, 72 milliseconds time to first audio at the 95th percentile of streaming requests, and 1.8 times Simba 3.2's throughput per serving instance. Those sit alongside the arena figures as vendor claims rather than as independent measurement, and they should be read that way.

What Sesame Preview is, precisely

Sesame Preview is a voice-first agent app, in public preview since May 2026, with named personas, memory that survives between sessions, and the ability to go and do things — search, notes, reminders — with a summary at the end of a call. It runs on iOS and Android, sells the three tiers above through the app stores rather than the web, and keeps a free browser Research Preview alive for users outside the app's countries at the same 30-minute daily ceiling as the Free plan.

Its engineering is oriented around turn-taking rather than rendering, and the company has published real work on it. TurnBench, released on 28 August 2026 with a paper on arXiv, measures the timing of conversation on thirty hours of dual-channel studio recordings — 154 dialogues, 106 voice actors in 53 pairs, recorded in separate sound-isolated booths — scoring end-of-turn detection and barge-in against two kinds of near-miss. The current board carries 25 ranked systems. None of them is a text-to-speech model, because the task is a different one.

And critically: there is no Sesame model endpoint. No id, no rate card, no reference-voice parameter. A developer cannot call Sesame Preview from code any more than they can call a phone call. The only loadable Sesame artefact is CSM-1B, and Sesame is explicit that CSM "can only model the text and speech content in a conversation — not the structure of the conversation itself."

The asymmetry that decides it

Simba 3.5 can do a narrow, partial version of what Sesame does. You can stream it, and streaming plus a good speech-recognition model is how a great many production voice agents are built today. What it cannot do is decide when to stop talking, listen while it speaks, or take an interruption — those are properties of the orchestration around it, and you would be building them.

Sesame cannot do any version of what Simba 3.5 does for a developer. You cannot put its voice in your product's UI at a per-character cost, render an audiobook with it, or localise a catalogue with it. You can talk to it, and your users can talk to it, and that is the entire surface.

So the decision reduces to a question about what you are shipping. If the voice is inside something you build, Simba 3.5 is the candidate and Sesame is not on the list. If what you want is the finished experience — a persona that holds a conversation and does errands — you are buying Sesame's app and the question of voice quality is being answered for you by a company that has not opened it to measurement.

For the localisation case, the gap is total

Simba 3.5 speaks 48 languages. Sesame Preview handles English, and its plans are indifferent to language because the product is English. For any workload where the script is not English, this comparison ends at the first paragraph of the spec sheets.

There is a nuance worth keeping, though. Speechify's 48 languages come with a caveat the vendor states plainly: a voice keeps the accent of its source recording, so a cloned American voice speaking Japanese sounds like an American speaking Japanese. The company's own guidance is to generate with a staff-pick voice native to each locale, which is why its voices endpoint now exposes a staff_pick_rank. Cross-lingual is supported; accent-free cross-lingual is not the claim.

Sesame's answer to accent is different in kind: a persona has a voice, and the voice does not travel because the conversation happens in one language. That is not a limitation for a consumer product and it is a hard stop for a catalogue.

Where OrcaRouter fits, and where it does not

OrcaRouter carries neither product. Simba 3.5 is reached through Speechify's own API; Sesame Preview is reached through an app store. Nothing here should be read as a claim otherwise.

What we do route is the smaller synthesis catalogue you would use as a control when testing a voice decision: openai/tts-1 at $15.00 per million characters and openai/tts-1-hd at $30.00, both at provider list price with no markup added. If you are weighing Simba 3.5 against anything — including against the OpenAI voice you already use — rendering the same script through both on one key is a cheaper experiment than a procurement cycle, and automatic failover means a provider incident does not take your render queue with it.

For the conversational case there is no router answer, and it would be dishonest to pretend otherwise. Sesame Preview is not an endpoint. If you need turn-taking in a product you are building, the comparison you actually want is Simba 3.5 plus your own orchestration against a realtime speech-to-speech API from a vendor that sells one — and that is a different article.

Which one, then

Choose Simba 3.5 if you are building: you need an endpoint, a per-character bill you can forecast, a countable language list, cloned voices that survive into 48 locales, and a top-of-board controlled-voice score earned on reference voices the vendor did not pick. It is also three weeks into a price schedule that has changed once, with a model-id migration — simba-multilingual and simba-english being served by Simba 3.5 — landing on 21 November, so pin your model id deliberately rather than relying on the default.

Choose Sesame Preview if you are using: you want a conversational agent with memory and task execution, you are comfortable that its voice is unmeasured by any independent board, and ten hours of talk time a day at $100 is a price you can attach to the value of the work it does. It is a finished thing rather than a raw material, and finished things are priced accordingly.

The mistake this comparison invites is treating them as competitors because both sound like people. They are sold in different units, consumed in different ways, and one of them cannot be called from code at all. That is not a tie to be broken — it is a fork in the road, and most teams already know which side of it they are standing on.