Gemini 3.8 Live vs Sesame Preview: The Voice You Can Admire vs the One You Can Call

Sesame's voice agent scores 4.5 out of 5 on "sounded human." Gemini 3.8 Live's scores 4.0. Sesame's also completes 71% of booking flows while Gemini 3.8 Live's completes 78%, and Sesame's makes function-call errors at roughly twice the rate. If you are choosing a voice stack for a business that books appointments, the model that sounds better loses — and it loses for reasons that have nothing to do with how good it sounds.

Those figures come from a third-party head-to-head run across 1,500 customer calls per model on booking scripts, and they carry two caveats that matter. They are not ours and not independently replicated, and the Gemini side used the early-2026 Gemini Live generation, not the model this article is about. Gemini 3.8 Live is Google's speech-to-speech release from September 15, 2026, available now on the Live API. Sesame Preview is Sesame's consumer voice assistant, free on iOS since May 2026 and broadly on Android since early August 2026, built around four named agents — Maya, Miles, Simone, Charlie — and widely described as the most human-sounding voice assistant anyone has spoken to.

The matchup is not really model against model. It is a product you can put in production against an experience you can only visit.

The shape of each thing

Gemini 3.8 Live is a developer surface. It ships as gemini-3.8-live and gemini-3.8-live-extended-thinking on the Live API, priced at $0.005 per minute of audio in and $0.018 per minute out, with text at $0.75 and $4.50 per million tokens and image or video input at $1.00 per million. It handles 97 languages with mid-conversation switching, processes visual input in near real time, runs tool and API calls in the background while the conversation continues, and watermarks every second of generated audio with SynthID. The Extended Thinking variant reasons and speaks at once, narrating progress so the line does not fall silent during a long task. Google's vendor-reported placement is second in the Speech Agent Arena, with the top spot on Artificial Analysis's Speech to Speech Quality Index at 82.6 for the family. Google's numbers, this week, unreproduced.

Sesame Preview is a consumer app. It is English-only. It has a 30-minute cap on a call, which drops to five minutes when you are logged out. Its agents hold context across sessions, search the web, and set reminders — but they do not book, buy, or file anything. There is no enterprise tier and no public pricing.

Both are full-duplex conversational systems with turn-taking good enough that people stop noticing it. That is where the similarity ends.

What the only head-to-head we have actually shows

The CallSphere comparison ran booking-flow scripts at volume and reported four numbers per model. On completion, Gemini Live finished 78% of bookings against Sesame Maya's 71%. On function-call errors, Gemini Live ran 3.3% against Maya's 6.7%. On p95 latency, Maya was faster — 480 ms against 640 ms. And on the subjective measure, "sounded human" rated 1 to 5, Maya took 4.5 against Gemini Live's 4.0.

Read those together and a pattern appears that no single row shows. Sesame wins the two rows a human notices and loses the two rows a business notices. Naturalness and latency are what a caller feels; completion rate and error rate are what shows up in the reconciliation. A voice that sounds like a person but drops one call in fifteen on a booking flow is not a better voice agent — it is a better demo.

Two honest caveats. The comparison is one vendor's blog, not a peer-reviewed evaluation, and 1,500 calls per model on a single script type is a narrow slice. And the Gemini figures describe the generation of Gemini Live shipping in early 2026, which is at least one full model behind 3.8 Live. Applying those numbers to Gemini 3.8 Live would be a guess. Nobody has published a head-to-head involving the 3.8 model.

The part almost every article gets wrong

Every write-up praising Sesame's voice is praising the app, and the app's voice is a larger, closed production model that Sesame has never released as an API. There is no way to buy it, license it, or call it. If you have read that Sesame's conversational timing is the best on the market and concluded that you can have it, the conclusion does not follow from the evidence.

What Sesame has actually open-sourced is CSM-1B — the 1B conversational speech model variant, released under Apache-2.0 in the Moshi family. It is a synthesis block, not an assistant: a Llama backbone predicts the first Mimi codebook and a smaller Llama decoder predicts the rest, which a Mimi codec turns into a 24 kHz waveform. It is English-only, it clones a voice from a clean 10-to-15-second 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. It is available hosted at $7 per million characters from a single third-party provider, roughly $0.35 an hour of audio, or free if you self-host.

The model card says the quiet part out loud: a fine-tuned variant of CSM powers the interactive voice demo, which is not the checkpoint you can download. So the honest version of "Sesame vs Gemini" is this: Sesame has a beautiful closed voice you cannot access, and a small open synthesis block you can, and they are not the same thing. Google has one model you can call and it is the one they advertise. That asymmetry is worth more to a builder than any naturalness score.

Spec contrast

• What you get — Gemini 3.8 Live a callable API today vs Sesame Preview a free consumer app with no API for its production voice

• Production voice licensable — Gemini 3.8 Live yes, pay per minute vs Sesame Preview no, never released

• Open artifact — Gemini 3.8 Live none, closed vs Sesame Preview CSM-1B, Apache-2.0, a different and weaker model

• Languages — Gemini 3.8 Live 97 vs Sesame Preview English only

• Tool calling — Gemini 3.8 Live background tool and API calls during conversation vs Sesame Preview agents remember context and search the web but do not execute tasks

• Session limit — Gemini 3.8 Live not published by Google vs Sesame Preview 30 minutes, 5 when logged out

• Naturalness — Gemini 3.8 Live unmeasured on this axis vs Sesame Preview rated best-in-class on listening impressions, with no independent benchmark of either

• Latency — Gemini 3.8 Live not published vs Sesame Preview 250–450 ms end to end for Maya, sub-300 ms streaming first byte, per third-party measurement

• Enterprise path — Gemini 3.8 Live private preview in Gemini Enterprise, coming soon to Customer Experience vs Sesame Preview none

• Stated direction — Gemini 3.8 Live Live API integrations with Agora, LiveKit, Pipecat, Fishjam and others vs Sesame Preview research partnerships and eyewear hardware planned for 2027

The naturalness row is the one people argue about, and notice what it says: neither model has an independent benchmark on how human it sounds. What exists is Sesame's app being praised by listeners and Google placing second in an arena. Those are not the same measurement, and the arena at least has a methodology.

What naturalness costs operationally

The mistake is treating naturalness as a feature you add. It is a trade you make, and the trade is usually against tool reliability — because the behaviours that make a voice sound human, the hesitations, the overlaps, the willingness to yield mid-sentence, are produced by a model that is spending capacity on prosody and turn-taking rather than on the strict grammar of a function call. That is the mechanism behind the 6.7% versus 3.3% error rate far more plausibly than any claim about one lab being better at engineering.

It also has an infrastructure cost nobody budgets. A voice that sounds human raises expectations, and a caller who believes they are talking to a person does not tolerate "I didn't catch that" twice in a row. Sesame's own 30-minute cap is a quiet acknowledgement that this class of model is expensive to run continuously — a cap Google does not publish for 3.8 Live, which is not the same as saying one does not exist.

Which one belongs in your stack

Choose Gemini 3.8 Live when the voice is a means to an end: scheduling, triage, support, anything where the caller wants an outcome. You get an API today, published rates, a tool loop that runs in the background, 97 languages, and an enterprise path if you need one.

Choose Sesame Preview when you are studying what conversational quality should feel like, or when you are building a consumer experience where the call itself is the product. It is free, it is genuinely the best-sounding thing available, and it is a fine place to calibrate your taste. Do not plan a production integration around it, because there is nothing to integrate with.

Most serious teams end up with a hybrid, and the hybrid is where routing stops being optional. Put a natural-sounding model on the opening and the acknowledgement turns, where latency and warmth are felt, and put a tool-reliable model on the transaction turns, where an error is a lost booking. OrcaRouter is built for that shape of stack — model fusion lets a panel of models answer together when a single model's judgment is not good enough, the routing DSL composes several into one call, and failover moves traffic off a model that starts misbehaving without a redeploy. Where a vendor does publish a rate, we pass provider list price through with no markup, so a price cut reaches your bill the same day.

What to watch: an API for Sesame's production voice, or a benchmark that measures naturalness the way the arena measures quality. Either one would turn this from 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 it faster than any score.

Most serious teams end up with a hybrid, and the hybrid is where routing stops being optional. Put a natural-sounding model on the opening and the acknowledgement turns, where latency and warmth are felt, and put a tool-reliable model on the transaction turns, where an error is a lost booking. OrcaRouter is built for that shape of stack - model fusion lets a panel of models answer together when a single model's judgment is not good enough, the routing DSL composes several into one call, and failover moves traffic off a model that starts misbehaving without a redeploy.