Phonic 1 vs Sesame Preview: The Dead-Air Problem, Solved Two Opposite Ways

Phonic 1 and Sesame Preview are both sold as conversational speech models, and they are not chasing the same job. Phonic 1 — the vendor spells it Phonic-1 — is a speech-to-speech endpoint at $0.07 a minute that became commercially available on October 29, 2026 and went straight to first place on Artificial Analysis's Speech to Speech Index at 87.1. Sesame Preview is the free companion app behind the Maya, Miles, Simone and Charlie voices: no public API, no rate card, no row on any independent leaderboard, and a stated plan to sell you glasses in 2027. Put the two names in one sentence and the first thing that falls out is that only one of them can be bought.

The second thing that falls out is more interesting. Both products are, underneath the marketing, answers to the same small humiliation: the half-second of silence while the machine decides what to say. Phonic attacks it by doing the work before the caller finishes. Sesame attacks it by getting better at knowing when the caller has finished. Those are genuinely different engineering bets, they are measured by instruments that do not overlap, and the choice between them follows from which bet your product depends on — not from a benchmark table.

What each one actually is, before the comparison gets slippery

Phonic 1 is an end-to-end speech model, audio in and audio out, with no transcribe-then-think-then-speak cascade in between. It shipped with TypeScript and Python SDKs, a CLI and a no-code agent builder, 51 documented languages with mid-conversation switching that keeps speaker identity, a documented intelligence_level setting that defaults to standard and can be raised to high, and enterprise terms that include SOC 2 Type II, a HIPAA BAA, zero data retention and a 99.9% uptime SLA. It is sold per minute in three tiers, and the company behind it is roughly ten people out of MIT and Stanford who raised a $27.5 million Series A led by First Round.

Sesame Preview is not a model you rent. It is a consumer product — an iOS and web app, free, rolled out across 39 countries from late May 2026, carrying four named agents with persistent memory, parallel web search that runs mid-sentence, note-taking, a texting mode and an incognito mode. Sesame's reported target for it is sub-300 ms first audio, and the company has been publicly candid that the model is still "too eager and often inappropriate" in tone and pacing. There is no enterprise tier and no way to license the production voice.

The one thing that does carry the Sesame name into a build is Sesame CSM-1B, published under Apache 2.0 on 13 March 2025. It is a base conversational speech model, English-first, built to generate Mimi audio tokens conditioned on prior turns. It is not the voice in the app: the Preview characters come from a curated fine-tune and a voice cast that were never released. Reviewers who have run both describe the open checkpoint as audibly weaker than the app, which makes "just self-host the Sesame model" a different product, not a licence-free shortcut.

Two attacks on the same half-second

Phonic's launch claim is the most concrete thing in this category: it says Phonic-1 is the only model that can act while the caller is still talking. In the demonstration the company published, the model fired all five tool calls mid-sentence and answered 1.1 seconds after the caller stopped; on the same script it puts GPT-Live-1 at 15.7 seconds. Read that as a demonstration of a design, not a benchmark — it is a vendor-selected task on a vendor-chosen script. But the design is the point. If the lookup is already done by the time the turn ends, the model never has to be brilliant at predicting the end of the turn; it just has to avoid getting in the way.

Sesame went the other way and made the prediction itself a research programme. In August 2026 it published TurnBench, an open benchmark for turn-taking dynamics: 30 hours of studio-recorded dual-channel dialogue, 154 conversations, 106 voice actors across six registers, triple-annotated at a Fleiss' kappa of 0.78, scored on recall at a capped false-positive rate with signed latency at the median. The headline finding is the one to keep: human speakers begin a turn transfer a median of 151 ms before the other person stops talking, and no system in the field matches that inside the false-positive budget. Semantic detectors are selective and slow; acoustic detectors are fast and noisy. Nothing is both.

That is the fork in the road, stated plainly. Sesame is trying to close the gap between 151 ms of human anticipation and the 368 ms to 1,234 ms the best detectors manage. Phonic is trying to make the gap irrelevant by removing the work that fills it. Both are legitimate. Only one of them is available to you as an endpoint today, and it is not the one that published the benchmark.

The board that scores one of them — and the one it sits out

Artificial Analysis's Speech to Speech Index is the only independent composite in the field, and Phonic 1 owns the top of it as of this writing. The rows that matter:

• Phonic 1 (High) — index 87.1, first; time to first audio 0.61 s; $4.20 per hour of input audio.

• Grok Voice 3.0 — 86.9; 0.81 s; $4.80 per hour.

• GPT-Live-1 (Astra, medium) — 83.2; 1.34 s; $5.83 per hour.

• Gemini 3.8 Live Extended Thinking (High) — 82.6; 1.35 s; $3.50 per hour.

• Eleven Realtime v1 Max — 77.3; 0.57 s; $4.80 per hour.

Note the last row. Eleven Realtime v1 Max is faster to first audio than Phonic 1 — 0.57 s against 0.61 s — and sits 9.8 index points lower. Time to first audio is not the contest Phonic won; it won the composite, which also weighs speech reasoning, agentic task behaviour, human preference and task success. Anyone quoting "fastest" off a single column is quoting the wrong column.

Now the asymmetry. There is no Sesame Preview row on that leaderboard, and there is no way for one to appear: the index runs models through an API harness, and Sesame Preview has no API. Sesame's own instruments do not fill the gap either. TurnBench is published by Sesame and, as of this writing, carries no Sesame entry — the board lists voice activity projection models, ESPnet, Kyutai's semantic VAD, Gemini 3.1 Live, OpenAI Realtime and Moshi, but not the company's own production voice. That is not a scandal; benchmark authorship and benchmark submission are different acts, and a companion app has no stable endpoint to point a scorer at. But it does mean the two headline numbers in this matchup — 87.1 on one side, sub-300 ms on the other — are not the same kind of claim. One was measured by a third party on a harness. The other is a target stated by the vendor about a product nobody outside the company has instrumented.

The column where Phonic 1 loses, on Phonic's own evidence

If the reason you are looking at Sesame Preview is that it sounds like a person, this comparison has an uncomfortable answer, and it comes from Phonic's own press material rather than from a critic.

Phonic commissioned blind pairwise listening tests from Podonos, a third-party listener-testing service, and reported the results in its launch post. Phonic-1 was preferred over GPT-Live-1 by 61.5% to 32.0%, over Grok Voice 3 by 66.5% to 26.5%, and over Eleven v4 Turbo by 78.0% to 14.0%. The headline the company chose to lead with is the one it lost: human voice actors were preferred over Phonic-1 by 57.0% to 32.0%, which the post describes as humans winning roughly 60% of the time, with substantial room still to improve. Those are vendor-commissioned, vendor-reported numbers and none has been independently replicated — but they are the vendor's numbers, and they say Phonic 1 is the best synthetic voice in its comparison set and still not a human.

Independent preference data points the same way. On the Speech Agent Arena that feeds the composite, Phonic 1 ranks eighth of 30 on listener preference at Elo 1,012 while ranking fourth on task success at 92.3%. The model that tops the index is not the model listeners most want to talk to. That is the same split we saw in Grok Voice 3.0 from the opposite side — xAI held the arena and lost the agentic column — and it is the structural reason a head-to-head between these two cannot be settled by one number.

Sesame's reputation, meanwhile, rests on exactly that column and on nothing more rigorous. The Maya and Miles demo went viral in February 2025 for breaths, hesitations, self-corrections and laughter, and the claim that circulated was that listeners could not reliably tell the voices from humans on short clips. That is an impression formed by an audience, not a scored test, and it has never been put in the same room as Phonic 1. There is no published experiment in which a listener hears Sesame Preview and Phonic 1 back to back. Anyone who tells you which one sounds better is telling you which one they have heard.

What a minute costs, and why the minute is the wrong unit

The price sheet is short and lopsided:

• Phonic 1 — $0.07 per minute on Starter with no minimum, 10 concurrent calls, 10 agents and 3 seats; $0.065 per minute on Pro against a $200 monthly minimum, roughly 3,100 minutes, with 50 concurrent calls; enterprise custom. Sign-ups before November 30 get $250 in credits, and customers migrating off a cascade get three months free.

• Sesame Preview — $0. Not per minute, not per seat, not per agent. Free in the app.

• Sesame CSM-1B — the weights are free, and the bill is your GPU hours plus the engineering time to assemble turn-taking, context conditioning and a language model around a checkpoint that cannot produce text on its own.

Run the productive comparison and the per-minute framing collapses. Ten thousand minutes a month on Phonic 1 Starter is $700, which is a rounding error for a call centre and a real budget for a side project. Sesame charges nothing because Sesame is not selling minutes — it is selling a daily habit ahead of a 2027 eyewear launch, and a habit is monetised per person, per year, on hardware, not per minute of audio. Comparing a rate card to a free app is comparing a supplier to a billboard.

Which leaves the honest version of the cost question: what does it cost to ship? For Phonic 1 the answer is one contract, one websocket, and ten concurrent calls on Starter unless you pay for Pro. For Sesame Preview the answer is that there isn't a number, because there isn't an offer. The nearest thing to a shippable Sesame is CSM-1B, and the nearest thing CSM-1B has to a rate card is a GPU invoice.

Three buyers, three different answers

You are moving call volume off a cascade and the calls have to complete. Phonic 1, and the case is the agentic column rather than the composite. Tool calls that fire before the caller stops are the difference between an agent that books the appointment and one that makes the caller repeat themselves. The $0.07 is cheap against the cost of a failed intake call, and the concurrency ceiling is the thing to size before you commit.

Your product is how it sounds. A companion, a coach, a tutor, a brand voice — anything where the user's reason for staying is the voice itself. Sesame Preview is the reference point and Phonic 1 is not yet the answer: on the vendor's own blind test, humans beat it nearly two to one. But the reference point is not purchasable, so the practical options are a licensed voice model with a measured preference score, or CSM-1B and a lot of your own engineering.

You want the split — Sesame-flavoured speech with an agentic backend. This is the combination both companies' roadmaps point at and neither sells. The buildable version is CSM-1B or another open speech model for the voice, and a conventional text model behind it for the thinking, wired together by you. That text stage is ordinary chat inference, which is the layer where one API for 200+ models with 0% markup actually applies: a single key for the reasoning leg, provider list price passed through unchanged, the routing DSL to send turn-taking chatter and policy-heavy reasoning to different models inside the same call, and automatic failover so a provider hiccup does not end the conversation. Neither Phonic 1 nor Sesame Preview is in that catalogue — Phonic 1 because it is a single-vendor endpoint and Sesame Preview because there is no endpoint to route — so this is a statement about the half of the stack you assemble yourself, not a claim about either model above.

What would change this answer

Three things, in ascending order of likelihood.

Phonic publishes listening-test numbers for the standard intelligence setting. Every score in this article, including the 87.1 and the 0.61 s, was produced at high. The API defaults to standard, documented as the lower-latency mode, and nobody has published what that configuration does to quality or to time to first audio. An integration that never touches the setting is running an unmeasured variant.

Sesame opens an endpoint, or puts its own production voice on its own turn-taking board. Either would turn this from a comparison of a product against a reputation into a comparison of two measurable things. Neither is on the roadmap Sesame has described, which runs through Android, then eyewear in 2027.

Someone runs both in the same room. A single blind test with Sesame Preview's production voices and Phonic 1 on the same prompts would settle the naturalness question in an afternoon, and it cannot be run by anyone who does not work at Sesame, because one of the two systems is not reachable by an outsider.

Until one of those happens, the defensible reading is this: Phonic 1 is the only one of the two you can deploy, it leads the only independent composite that scores either of them, it wins on the work-done-before-the-turn-ends bet, and it is measurably not yet a human voice. Sesame Preview is the better-sounding thing you cannot have, from a company that is not currently in the business of selling it. Those are not two options on one menu. They are a supplier and a demo, and the faster you stop comparing their spec sheets, the faster you can decide which one your product actually needed.