GPT-Live-1 vs Cartesia Sonic 3.6: 1위 TTS 리더보드는 인터럽션을 측정하지 않는다

Cartesia Sonic 3.6 currently leads both of Artificial Analysis's text-to-speech arenas — 1,275 Elo on the Provider Voice board and first on the Controlled Voice board as well, ahead of every commercial engine with a bigger platform around it. GPT-Live-1 is not ranked on either, because it is not a text-to-speech model. It sits joint sixth of fourteen on a different Artificial Analysis board, the Speech to Speech Index, at 70.3%. Read together, those two facts describe the actual split between these products: Sonic 3.6 is very likely the better voice, and GPT-Live-1 is the only one of the two that can hear the caller start talking and stop.

Both are live and both are commercially available — Sonic 3.6 on Cartesia's own API since its stable snapshot landed on August 27, 2026, GPT-Live-1 on Open​AI's developer API since September 10, 2026, at $0.05 per voice minute. Choosing between them is not a quality decision. It is a decision about which half of a voice agent you are buying.

Two arenas, two questions, and why the split is real

Artificial Analysis runs its speech boards in a way that is worth understanding before either Elo is quoted at you. The Provider Voice arena ranks engines using each vendor's own native voices — it measures the voice you actually get out of the box, which is the number a buyer cares about. The Controlled Voice arena clones every model onto the same eight reference voices so that listeners compare the synthesis engine rather than the voice talent. Sonic 3.6 leads both, which is rarer than it sounds: the two boards have had different champions for most of 2026.

Neither board contains a single sample of a model being interrupted. They measure how speech sounds, in isolation, to a listener. The Speech to Speech Index is built differently — it weighs speech reasoning, agentic performance, arena preference and task success into one figure, and it only admits models that are natively speech-to-speech. Cartesia has no entry there. Not because Sonic 3.6 is bad, but because a text-to-speech engine has nothing to submit.

So the 1,275 and the 70.3% are not competing numbers, and any comparison that puts them on one line is quietly lying to you. What they jointly establish is that quality has stopped being the axis on which this decision turns.

차원별로

• Kind — GPT-Live-1: full-duplex speech-to-speech, one model for the whole loop vs Cartesia Sonic 3.6: streaming text-to-speech, paired with Ink-2 speech recognition for agents

• Independent quality — GPT-Live-1: 70.3% on the Speech to Speech Index, joint sixth of fourteen vs Cartesia Sonic 3.6: 1,275 Elo on the Provider Voice board, from 1,755 samples, and first on the Controlled Voice board as well

• Interruption handling — GPT-Live-1: a property of the model, which decides many times a second whether to yield the turn vs Cartesia Sonic 3.6: an integration pattern, where the client cancels the synthesis context and relies on word-level timestamps to cut at the right moment

• Price — GPT-Live-1: $0.05 per voice minute, billed per second, reasoning backend metered separately vs Cartesia Sonic 3.6: about $49 per million characters under credit plans, plus $0.06 per minute of agent call duration and $0.014 per minute for a Cartesia phone number

• Latency — GPT-Live-1: no model-level figure published; turn-taking latency given as 0.798 seconds in Open​AI's own testing vs Cartesia Sonic 3.6: under 90 milliseconds time-to-first-audio, vendor-stated and not independently measured end to end

• Throughput — GPT-Live-1: concurrent sessions, capped at 25 on tier 1 and 500 on tier 5, with no free tier vs Cartesia Sonic 3.6: about 132 characters per second, roughly double ElevenLabs' v3 rate on the same comparison, with a free tier of 20,000 monthly credits

• Languages — GPT-Live-1: uneven fluency outside the major languages in launch coverage vs Cartesia Sonic 3.6: 44 languages across 61 locales, with Odia and Urdu added in this release

• Voice control — GPT-Live-1: twelve presets, tone and pace through prompting, no cloning vs Cartesia Sonic 3.6: more than 500 preset voices, instant cloning, professional cloning on higher tiers, word and phoneme timestamps, and no SSML — the model adapts pacing from the text itself

Barge-in is a feature of the architecture, not a checkbox

The single most common thing buyers get wrong about this comparison is treating interruption support as a boolean. Cartesia's documentation is candid about how its version works: when the caller talks over the agent, the client sends a cancellation for the current synthesis context, and the word-level timestamps returned on the WebSocket connection are what let you stop playback on the correct syllable. That is a competent design and it is how most production voice agents handle barge-in today.

It is still a design in which the decision to stop talking was made by your application, using a voice-activity signal, at whatever latency the round trip allows. GPT-Live-1 moves that decision inside the model. It is continuously deciding whether to keep listening, pause, take the turn or hand it over, which is why Open​AI's own figures report 80.1% interactivity against 45.4% for GPT-Realtime-2.1 on Full Duplex Bench v1.5, with turn-taking latency of 0.798 seconds against 1.41. Those numbers are Open​AI's, published with the release and not reproduced by anyone outside the company — but the architectural difference behind them is not in dispute, and it is the one thing a cascade cannot approximate by tuning.

Where the cascade wins is everything downstream of the sentence. Cartesia's sub-90-millisecond time-to-first-audio is a vendor figure and it measures the model, not the conversation: the number a caller experiences also includes speech recognition, turn detection, the language model's generation time, network transit and audio buffering. That is exactly why a cascade is worth building when you need to control each stage — a fast small model for simple turns, a frontier model for the hard ones, and a synthesiser that never makes you wait.

That middle stage is the only part of either stack that is genuinely portable, and it is the part that decides both the quality and the cost of the call. Roughly 190 models from eleven upstream providers sit behind one OrcaRouter key at provider list price with zero markup, and the routing DSL lets a single endpoint send straightforward turns to a cheap model and escalate the complicated ones to a frontier model — the pattern a voice agent actually wants, applied to the stage that varies most. Neither Sonic 3.6 nor GPT-Live-1 is reachable through a routing layer; the voice layers belong to their vendors.

Running the money

The two pricing models do not line up, so the conversion is worth doing properly. Speech runs at roughly 150 words per minute, and English averages about five and a half characters per word, so a minute of spoken output is around 800 to 900 characters. At $49 per million characters, Cartesia's synthesis costs about four cents per minute of audio.

Then the rest of the cascade arrives. Cartesia charges $0.06 per minute of voice-agent call duration and $0.014 per minute if you use its phone number, on top of the characters. Add speech recognition and a language model for the text and you are past a dime a minute before anyone has said anything difficult. GPT-Live-1's $0.05 per voice minute covers the listening, the turn-taking and the speaking, with the delegated reasoning billed separately at the backend model's normal rate — which for a reservation call with a search or two attached is a real number, not a rounding error.

The crossover sits at volume and at difficulty. Short, repetitive, high-volume calls — confirmations, notifications, simple lookups — favour the cascade, because a cheap model answering a scripted question is the cheapest thing in this comparison. Calls where the caller goes off-script, speaks over the agent, or changes their mind mid-sentence favour the end-to-end model, because the cascade's failure mode there is not a wrong answer, it is an awkward one.

어느 걸 고를까?

Choose Cartesia Sonic 3.6 if you want the best-scoring voice available and you are willing to own the conversation logic yourself. It is the right call for narration, for high-volume scripted agents, for teams with an existing speech recognition pipeline, and for anyone who needs the word-level timestamps that make precise interruption handling possible. You get a free tier, more than 500 voices, cloning, 44 languages and the top of both quality boards, and you get to keep control of every stage.

Choose GPT-Live-1 if the interruption is the product. Anything where being talked over is the normal case rather than an edge case, and where a 0.8-second turn handoff beats a 90-millisecond start on a sentence nobody was finished asking. You accept a closed, hosted, per-minute model with twelve voices and no cloning, and you accept that its independent quality score is mid-table.

The temptation is to read 1,275 against 70.3% and call it settled. It is not, and the difference between those two numbers is the whole argument: one measures how a voice sounds, the other measures whether it is worth talking to.