Eleven Realtime v1 Max vs Sesame Preview: The Voice That Wins the Ear Loses the Task

Ask someone which voice assistant sounds most human and you will hear S​esame Preview. Ask which model an enterprise can put behind a phone number and the answer is Eleven Realtime v1 Max, a speech-to-speech model announced October 20, 2026, that resolves 87.1% of tasks on a fixed evaluation, answers in 0.57 seconds and is available through ElevenAgents or an O​penAI-compatible Realtime API. S​esame Preview is a consumer app with four named companions and no way to buy the voice inside it. The documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.

That is not a knockout, because the two products are chasing different things and the case for S​esame's voice is real. S​esame's researchers were early and public about the idea that conversational timing — the mm-hmm, the overlap, the willingness to yield mid-sentence — is the uncanny valley, not the timbre. E​levenLabs' documentation makes the same bet from the other direction: Mini is described as detecting whether a speaker has finished or is merely interjecting. Two labs, one thesis, opposite business models.

Where the comparison gets interesting is what each one has actually measured. E​levenLabs publishes six numbers and an evaluation partner. S​esame publishes an experience.

What each one is, precisely

Eleven Realtime v1 Max is a single network that performs recognition, turn-taking, reasoning and voice generation together, positioned for complex tasks with multi-step tool use: customer support, lead qualification, assisted scheduling. Its cheap sibling, Eleven Realtime v1 Mini, is the default for new ElevenAgents agents and targets natural conversation at lower cost. Max is reachable two ways — from the ElevenAgents dropdown, where E​levenLabs manages the LLM, T​TS model, ASR provider and reasoning settings for you, or through a Realtime API that speaks O​penAI's Realtime event protocol over WebSocket at wss://api.elevenlabs.io/v1/convai/realtime.

S​esame Preview is a free consumer voice assistant: four named agents — Maya, Miles, Simone, Charlie — available on iOS and the web, described by nearly everyone who tries it as the most natural voice assistant on the market. It holds context across sessions, searches the web and sets reminders. It does not book, buy or file anything. There is no enterprise tier, no published pricing and no public API for the voice you hear in the app.

Both are full-duplex conversational systems with turn-taking good enough that people stop noticing it. That is where the similarity ends and where the "which is better" question stops having a single answer.

The numbers one of them has

Artificial Analysis evaluated Eleven Realtime v1 Max independently on the same harness it applies to every speech-to-speech model. These are its figures, not E​levenLabs':

• Speech reasoning (Big Bench Audio) — Max 98.8%

• Conversational dynamics (Full-Duplex-Bench) — Max 94.1%

• Task success rate — Max 87.1%

• Arena preference (Elo) — Max 997

• Time to first audio — Max 0.57s

• Cost per hour of input audio — Max $4.80

• Agentic performance (τ-Voice) — Max not measured

• Speech to Speech Index — Max not scored

S​esame Preview appears on none of these rows. It has no public API, so no independent harness can run it, and the app's own behaviour has never been benchmarked on task completion or latency by anyone outside S​esame. When people say S​esame sounds better, they are reporting a listening impression, and it is a legitimate one — it is simply not the same measurement as an arena with a methodology and a preference model behind it.

That cuts both ways. The 997 arena Elo is a real number from real listeners, and it puts Max behind G​rok Voice 3.0's 1,064, G​PT-Live-1 (Sol, low)'s 1,047 and G​emini 3.8 Live's 1,078. E​levenLabs wins the reasoning and speed rows and does not win the preference row. If your instinct is that S​esame's voice beats Max's, the arena result is consistent with that instinct — it just happened to be measured against other API models, not against S​esame.

The open-weight question, which is where most comparisons go wrong

Every write-up praising S​esame's voice is praising the app, and the app's voice is a closed production model that S​esame has never released as an endpoint. What S​esame has open-sourced is CSM-1B, a 1B conversational speech model in the Moshi family, under Apache-2.0. It is a synthesis block, not an assistant: a Llama backbone predicts the first Mimi codebook, a smaller Llama decoder predicts the rest, and a Mimi codec turns that into a 24 kHz waveform. It is English-first, it clones a voice from a short 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 S​esame's own model card supports that reading: a fine-tuned variant of CSM powers the interactive demo, and that variant is not the checkpoint you can download.

So the honest version of "S​esame vs E​levenLabs" is this. S​esame has a beautiful closed voice you cannot call, and a small open synthesis block you can, and they are not the same model. E​levenLabs has a callable model whose voice is measurably good but not the best-liked, and a price. The asymmetry that matters to a builder is not which sounds better; it is which one has an endpoint, and that is not close.

Where the merge shows up in production

E​levenLabs' argument is that merging recognition, turn-taking, reasoning and voice into one tuned system removes the seams a cascade leaves.

In ElevenAgents the merge is literal and it has a cost: while a Realtime model is set, the LLM, reasoning effort, thinking budget, parallel tool calls, cascade timeout, T​TS model, expressive mode and ASR provider are all managed by E​levenLabs and ignored on the agent. A workflow node cannot override them, and the agent cannot use a custom LLM or Amazon Bedrock at all. Choose the model and you choose the whole stack beneath it.

The compliance matrix is uneven in a way that will surprise people rolling this out across regions:

• Data residency — Max runs in the US and EU but not India or Singapore, where only Mini is offered

• Zero Retention Mode and HIPAA — Max is supported in the US and EU; Mini is not supported under either

• Token limit — Max requires 1,024 tokens or no limit, and workflow nodes cannot override it

• Agent transfers — a Realtime-mode agent can only transfer to and from other Realtime-mode agents, though Mini-to-Max is allowed

• Transport — no WebRTC at launch; the Realtime API is WebSocket only

Those rules are enforced at publish, not at save: a draft that breaks one saves fine and fails when you publish, with an error naming the setting. If you are prototyping, that is a minor annoyance. If you are shipping to a regulated queue, it decides which variant you can use in which environment, and the cheaper one is not always the compliant one.

Which one belongs in your stack

Choose Eleven Realtime v1 Max when the voice is a means to an end and the call has to produce an outcome — support triage, scheduling, qualification. You get an endpoint today, an O​penAI-compatible event protocol that makes migration from a Realtime integration a two-string change, a measured 87.1% task-success rate, the fastest strong reasoner on the board, and a price you can model at $4.80 per hour of audio. Accept that it is still rolling out, that its agentic behaviour is unmeasured, and that its voice does not top the preference arena.

Choose S​esame Preview when you are studying what conversational quality should feel like, or building a consumer product where the conversation is the product. It is free, it is the best-sounding thing most people have tried, and it is a fine way to calibrate your taste. Do not plan an integration around it: there is nothing to integrate with, and CSM-1B is not the model in the app.

Choose CSM-1B only if you want an open, self-hostable synthesis block and are prepared to build the rest — an ASR, a language model, a turn-taking layer — around it. It is Apache-2.0 and it is genuinely useful in that role. It is not a competitor to Max and was never meant to be.

The hybrid, and the half a router can help with

Most serious voice products end up mixing models rather than picking one. A natural-sounding model on the opening and acknowledgement turns, where warmth and latency are felt; a task-reliable model on the transaction turns, where an error is a lost booking. Neither Eleven Realtime nor S​esame makes that easy inside a single connection — E​levenLabs' Realtime API is a separate product from ElevenAgents and runs over WebSocket only, and S​esame has no API to mix with.

Where a routing layer does earn its place is one layer down, on the work the voice model delegates. On OrcaRouter that is 209 models from 16 upstream providers behind one key, at provider list price with zero markup, with automatic failover when a provider errors. Model fusion lets a panel answer together when one model's judgment is not enough; the routing DSL composes several into one call. Eleven Realtime v1 Max itself is not on our catalogue — there is no third-party endpoint for it — so the honest boundary is that we can route the retrieval, the summarisation and the post-call analysis, not the voice.

What would change this article is a benchmark that measures naturalness the way the arena measures preference, run across both S​esame's app and the API models — or an endpoint for S​esame's production voice. Either one would turn a choice between access and experience into an actual comparison. Until then, one of these models you can call and the other you can only admire, and that settles more than any listening test.