Title card for 'Sesame Preview Arrives on Android': large title, subtitle 'The voice agent reviewers call one of the best voice modes on the market — free, with no API yet', and two flat icon cards — 'Voice-first agents' with a waveform-and-person line icon and a badge reading 'Maya · Miles · Simone · Charlie', and 'On iOS and Android' with a phone line icon and a badge reading 'free — no API yet'. Footer: 'Personal agents for curious people — Sesame, August 2026.' OrcaRouter logo composited bottom-right.
Guides & Insights

Sesame Preview Now Has Plans: Free, Plus at $20, Pro at $100 — and Still No Voice API

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Nearly six weeks after the October upgrade that gave its agents a computer, Sesame Preview — an app in public preview since May 2026 — has a price list. Sesame's own plans page now sells three tiers — Free at $0 with 30 minutes of talk time a day, Plus at $20 a month on four times that allowance, and Pro at $100 a month on twenty times — billed through the App Store and G​oogle Play rather than on the web. Those are vendor-stated prices on the vendor's own page: not a measured number, not an independent review. Everything underneath them is unchanged. The production voice still has no API, no model ID and no independent score, and the only Sesame model anyone outside the app can load remains the open-weight CSM-1B.

The three tiers, as published:

• Free — $0. 30 minutes of talk time a day, plus a limited weekly credit allowance for texts and tasks.

• Plus — $20 a month. Two hours of talk time a day, four times Free's weekly credits, and a one-week free trial that anyone can use once.

• Pro — $100 a month. Ten hours of talk time a day and twenty times Free's weekly credits.

Sesame draws the line that matters itself: talking on a call spends talk time, and "anything the agent goes and does" spends credits, whether you asked by voice or by text. The two meters therefore fail differently. Talk time resets daily and credits reset weekly. Running out of credits does not cut the call — it drops you from Max intelligence to Standard, and scheduled skills pause until the weekly reset, then resume where they left off. Running out of talk time is the harder wall: you can still text, but model intelligence drops and some capabilities are restricted. On the tier list, "four times the talk time" and "four times the credits" are two separate promises that can run out on different days.

The number that actually decides the tier is not the headline price but the add-on rate. Sesame sells extra talk time at $5 for one hour and $20 for five, and credits at $5 per thousand and $20 per five thousand. Packs never expire and are spent only after your plan allowance is gone. At that rate, Plus's two hours a day — roughly 60 hours in a 30-day month — costs $20, while $20 of à-la-carte talk time buys five hours. If you will use the assistant more than a few times a week, the subscription is about twelve times the value of the packs, which is presumably the intent: Sesame's own FAQ calls every plan "a much better deal than buying the equivalent amount of additional time or credits."

Two structural details matter more than they look. First, nothing is sold on the web: purchases run through the App Store or G​oogle Play, so the browser Research Preview cannot be upgraded at all. Second, and this is where "the preview is ending" reads it slightly wrong, the web preview is not going away. Sesame's FAQ says users outside the app's countries "can keep using the Research Preview for free, with a daily talk time limit that matches the Free Plan: 30 minutes a day," and that talk time and credits are tied to the account, so the same daily ceiling applies to web calls. What ended is the open-ended allowance. The preview itself survives as the free tier.

The app itself keeps shipping alongside the new billing. The App Store lists version 48.0, dated November 13, with release notes advertising "better standard and Max Intelligence"; G​oogle Play shows the app updated November 9 at 50,000+ downloads and a 4.9-star average across about 2,000 reviews. The screenshot below is not that build — it is the August capture this article originally ran, of a page that read "Preview available now on iOS and Android" and named no price at all. It is kept because it is the last state of the product before the tiers existed.

Screenshot of Sesame's official Mobile Preview page (sesame.com/mobile-preview), showing 'Personal agents for curious people', 'Preview available now on iOS and Android', and store links for App Store and Google Play. Captured August 6, 2026.

What Sesame actually is

Sesame is not a model company that happens to have an app; it is a model company whose app is the point. The lab, founded in 2023 by Brendan Iribe, Ankit Kumar and Ryan Brown and backed by a16z, Sequoia, Spark and Matrix, built the conversational speech model (CSM) behind the "Maya" demo — a voice with breath, hesitation and interruption timing so convincing that first-time listeners assumed a person was on the line. The product it sells now is "personal agents for curious people". October added a destination: on October 6 the company's chief product officer, Nate Mitchell, framed the assistant as the operating system for a pair of all-day glasses — "a voice OS on the frames," in the company's phrasing — and a day earlier Sesame announced that the frames will be made in Japan and arrive in 2027. The assistant is the product; the glasses are the reason it exists. The plans page added the part that turns it into a business.

The four agents, and what the app actually does

The four personalities are unchanged and still the most distinctive thing about the product: Maya (a witty collaborator), Miles (a laid-back conversationalist who tells it like it is), Simone (a dynamic thought partner for deep dives) and Charlie (a casual sounding board). Each keeps its own memory and, as of October, its own app access — you can give Maya your email and keep Charlie out of it. On that base the assistants search the web, run deep dives, take notes and reminders, send a summary after a call, and answer in text when you cannot talk out loud; an Incognito Mode reads past memories without writing new ones. What the plans changed is not what the app does but how much of it you get: every feature is listed as available on every tier, and the difference between $0, $20 and $100 is allowance, not capability.

What the October release actually added

Start with the voice, because everything else sits on it. Sesame says it "updated our Conversational Speech Model to improve expressivity and pronunciation," and that the change "makes a huge difference, especially over longer conversations." The same post says the characters' personalities now "come through more clearly", with less sycophancy and fewer repeated phrases. That is a vendor claim with no numbers attached: no model ID published, no latency figure, no benchmark, and the app's voice is still not an endpoint anyone can measure. Treat it as a directional statement from the only party that has heard both versions.

The larger change is that the agents now have a body. "With the October release, each agent is now smarter, with access to its own computer," Sesame writes — meaning an assistant can browse and act rather than only talk. The demonstration the company offers is pointed: its own team says it has been using Maya, Miles, Simone and Charlie to "orchestrate other agents like Devin, Claude Code, and Codex", over voice, on walks and commutes. If that works as described, Sesame has stopped competing only with voice assistants and started competing for the control layer above coding agents. It is also the claim in this release a reader can most easily test personally — and the one no outside reviewer had published a measurement of when this was written.

The plumbing that makes it useful is new too. Sesame's getting-started guide, updated October 5, lists G​oogle apps — Gmail, G​oogle Calendar, G​oogle Drive — as connectors, some tagged Beta, plus support for external MCP servers over OAuth, with a frank warning to trust the developer before connecting one. On top of the connectors sits the "skill": a plain-language playbook you build in conversation and then run on demand or on a schedule, with or without connected apps. Image generation arrived as one of the agent's own tools: the app lists "Image by Sesame" among them, described as generating, editing and analyzing images, and the assistant will offer to generate one from a description or to tell you what is in a photo you send. Note the provenance there — that is the in-app tool list, not something Sesame's own announcement says, and neither the app nor the company names the image model behind it.

The plans page adds one detail the October notes did not spell out: the app now has two intelligence levels, Standard and Max, and Max is the one that spends credits as well as talk time. That makes the credit allowance a quality ceiling rather than a task quota — a distinction the store listing supports, with App Store release notes for version 48.0 promising "better standard and Max Intelligence."

The limits that did not change

There is still no API and no model ID for the voice, so none of this is something a developer can call — and the plans page does not change that. Every figure on it is a consumer subscription; none of it is a per-token, per-minute or per-character rate for the model. The image tool will not tell you which model powers it. How reliably an agent orchestrates Devin or Claude Code unattended is an open question the vendor has answered only with its own anecdote. English is still the only officially supported language — the guide notes the assistants also handle French, Italian, German, Spanish, Chinese, Japanese and Korean "early", which is a warning rather than a feature list. You must be at least 18. And app access is per-assistant, which is good for privacy and awkward in practice: what you connect for Maya, Charlie cannot see. Sesame still says it runs no ads and does not sell data, a claim the paid tiers now make easier to keep.

The voice is the product — and it is unmeasurable so far

What makes Sesame worth writing about is still not a benchmark score, because there is still not one for the app. The production model is unreleased: no model ID, no API, no latency target, and no listing on Artificial Analysis' speech-to-speech leaderboard the way callable voice models are. What you can verify is the architecture family. Sesame's CSM is a two-decoder design — a Llama-backbone predicting the first audio codebook and a small "depth" decoder predicting the rest — over Kyutai's Mimi codec, producing 24 kHz mono speech, and the open-weight CSM-1B (Apache-2.0) is the one piece of it anyone can download. The October notes are the first confirmation in months that the production CSM itself has moved, and the company still published nothing you can point a benchmark at. One instrument does exist, and it measures something adjacent: on August 28, 2026 Sesame published TurnBench, an open benchmark for turn-taking in spoken dialogue — 30 hours of triple-annotated dyadic speech, a scoring protocol drawn from conversation analysis, a public leaderboard and results for 14 systems, built with Carnegie Mellon, Mundo AI and Oto. It is genuinely open, it is Sesame's own, and it scores turn-taking rather than this app's voice.

The scoreboard below is that older August snapshot. It is now short in two directions. It has nothing to say about computer use, connectors or image generation, because none of them existed when it was drawn; and its price row still reads "free (preview, no paid tier)", which stopped being true the week the plans page went up. The title card at the top of this page is from the same August batch and carries the same outdated "free — no API yet" badge. Read both as a record of what the app was, not of what it now costs.

Single-column scoreboard for 'Sesame Preview — the scoreboard': six rows reading 'Platforms: iOS + Android + Web', 'Price: free (preview, no paid tier)', 'Agents: Maya · Miles · Simone · Charlie', 'Independent score: none yet', 'Open model: CSM-1B, $7/M chars', 'Languages: English only'. Footer: 'Sesame app model unreleased; CSM-1B per OpenRouter; no independent voice benchmark exists.' OrcaRouter logo composited bottom-right.

What an hour of Sesame actually costs

Priced per hour of talking, Plus works out to roughly 33 cents. That is worth putting next to the API market, because it is the only figure Sesame now publishes that can be compared to anything. G​rok Voice Think Fast 2.0 charges $0.08 per audio minute, which is $4.80 an hour. GPT-Realtime-2.1 is billed per audio token at $32 per million input and $64 per million output, a unit that cannot be converted into hours without knowing how much you say.

That comparison is not apples to apples and should not be read as one. G​rok Voice is an API you build a product on; Sesame is an app you talk to. Sesame's credits meter the agent's actions on top of the talk time, so a heavy week of tasks can cost allowance that the hourly figure never sees. And nothing independent has measured whether the two voices are in the same league — the one number that would settle it does not exist for the app. What the 33 cents does say is that Sesame has priced a flat consumer subscription against a market that bills by the minute, and that at $20 a month it is selling hours well below what a metered voice API charges for them. Whether that is generosity or a launch discount is a question the next pricing page will answer.

Where Sesame Preview sits in the voice landscape

In 2026 the realtime-voice field splits into apps and APIs, and October added a third axis: who controls the agents. On the app side, Sesame Preview competes for the same attention as the voice modes inside ChatGPT and G​emini Live — with one new difference, since those voice modes come bundled into a subscription that also buys a text model, while Sesame's $20 buys only the assistant. That makes the free tier the real contest: 30 minutes a day is a demo length, and the moment a reader passes it the decision becomes whether this voice is worth a standalone $20. On the API side, the callable options are the models you can point a WebSocket at: G​rok Voice Think Fast 2.0 at $0.08 per audio minute with a 0.70-second time-to-first-audio, ranked first on the τ-Voice agent benchmark and second on Artificial Analysis' speech-to-speech quality index at 82.9% (those figures are Artificial Analysis', not x​AI's), or GPT-Realtime-2.1 at $32 per million audio input tokens and $64 per million audio output tokens. The third axis is the one Sesame moved onto in October: not "call our voice model" but "talk to our agents and they will drive your other agents." That is a claim about orchestration, and nobody measures it — the vendor's description is the whole of the evidence so far.

The chart below is the API-side board as it stood on August 6, 2026: a reminder of what Sesame competes for attention with, and of the kind of figure the app's voice still does not have.

Screenshot of the Artificial Analysis speech-to-speech leaderboard (artificialanalysis.ai/speech-to-speech, captured August 6, 2026), where callable voice models such as Grok Voice Think Fast 2.0 carry index scores and time-to-first-audio figures — the API-side voice market that Sesame Preview, an app with no API, currently has no equivalent for.

What it means for builders

The consumer side now has a price; the developer side still has none. If you want to build a product on Sesame's production voice today, you cannot — there is no API for it, and no plan on the pricing page is a developer plan. The only Sesame model a developer can call is the open CSM-1B base, which you would self-host or reach through a provider that lists it: 1B parameters, 4K context, a text-to-speech model with zero-shot voice cloning but no tool-calling of its own. What October changes is the direction of travel: the assistants connect to MCP servers, and Sesame says it is "adding developer tooling so you can make your own Sesame apps later this year." What the plans add is a hint of the business model that tooling will sit on — metered talk time plus metered agent work, sold as tiers. An assistant you can extend and now pay for by the hour is one step from an assistant you can call by the token.

The more common 2026 way to assemble a custom voice agent remains pairing a speech layer with a text brain that decides what to say — and that brain is exactly what a router fronts. OrcaRouter gives you 200+ models behind one OpenAI-compatible endpoint at provider list price, passed through with zero markup, so when a vendor cuts a price it is live here the same day; automatic failover is the safe way to put a just-released, unproven model on a real path beside a proven one. To be straight about our own coverage: no Sesame model is a route here today, CSM-1B included — so none of the above is you renting Sesame's voice. It is the pattern for the text side while the voice stays closed, and the pattern you would want on the day it does not.

The "drive mode" test

TestingCatalog's phrase from August — "my drive mode default" — is still the sharpest summary of what Sesame Preview is for, and October made it more literal and more consequential. Drive mode is the hardest case for a voice assistant: hands busy, eyes on the road, no screen to fall back on, no patience for a turn-taking failure at speed. Add computer use and the stakes change — you are now delegating tasks you cannot verify from the driver's seat, to an agent that will go and act and come back with an update. Expressive prosody that carries meaning without a display, and interruption handling that lets you cut in mid-sentence, are still the reasons this voice is worth the drive.

The plans add a less romantic failure mode to that test. Because task work spends credits rather than talk time, the thing that runs out in a car is not the conversation but the doing: hit the weekly credit ceiling and calls continue at Standard intelligence while any scheduled skill pauses until the reset. Fifteen minutes of talk time left and no credits is a worse position than ten minutes and a full allowance, and the app warns you before you get there rather than after.

What to watch next

Five things would change this story. First, an API or model ID for the production voice — the event that turns Sesame from a consumer app into something a developer can route to, and the moment it can be priced per token instead of per month. Second, whether the plan structure holds its shape: today every feature is advertised on every tier, so the tiers differ only in allowance. A single feature moving behind Plus would tell you the metering was a test rather than a design. Third, the developer tooling Sesame has promised for "later this year": if you can build a Sesame app, the app's voice becomes programmatically interesting even without an API in the classic sense. Fourth, open weights — CSM-1B has been alone on Hugging Face since March 2025, and a newer or larger open CSM would let anyone run the voice themselves instead of renting hours. Fifth, the eyewear, which is no longer a rumour: across October 5 and 6 the company dated the glasses to 2027, said they will be made in Japan, and described the assistant as the operating system they run.

FAQ

Is Sesame still free?

Partly. The app downloads free and the Free tier keeps full access to every feature, capped at 30 minutes of talk time a day and a limited weekly credit allowance; the browser Research Preview runs on the same 30-minute daily ceiling. Plus is $20 a month for two hours a day and four times the credits, Pro is $100 a month for ten hours a day and twenty times the credits, and à-la-carte packs ($5 an hour of talk time, $5 per thousand credits) let you top up without subscribing. Everything is billed through the App Store or G​oogle Play.

Does Sesame have an API for the app's voice?

No. The production model behind the app is unreleased — no model ID, endpoint or per-token pricing — even though the October release says the underlying Conversational Speech Model was updated. The paid plans are consumer subscriptions, not developer rates. The only Sesame model a developer can call today is the open-weight CSM-1B base, which you would self-host or reach through a provider that serves it. The assistants can connect to external MCP servers, but that lets them use your tools; it does not let your code use them.

How good is the voice, really?

There is no independent score. The app model is not listed on Artificial Analysis' speech-to-speech leaderboard, and Sesame has published no third-party benchmark of it. What exists is the vendor's own account of what changed in October — more expressivity, better pronunciation, less sycophancy — plus Sesame's open TurnBench, which measures turn-taking rather than this app, and the open CSM-1B, whose naturalness is the one part of Sesame's voice work anyone can independently run.

Sesame Preview spent October becoming a product and has since become a business, and the two facts are worth keeping apart. As a user, the question is no longer whether the voice is impressive — it is whether 30 minutes a day covers how you actually talk, whether $20 buys enough room, and whether the weekly credit ceiling lands in the middle of something you scheduled. As a developer, nothing changed at all: no API, no model ID, no per-minute rate, no score. Try the voice on the free tier before you subscribe, watch what the credit cliff does to a real week rather than a demo one, and keep one eye on the model ID — because the moment Sesame ships one, this stops being a consumer story and becomes a developer story overnight.