Fish Audio Live 3 vs GPT-Live-1: Both Are Half a Product, and the Meter Says Which Half

Fish Audio Live 3 and GPT-Live-1 are both full-duplex voice models, both argue that turn-taking belongs inside the speech model, and both are, on their own, incomplete. GPT-Live-1 listens and speaks and then hands the thinking to a separate backend model — that is its design, not a limitation — and it bills $0.05 per voice minute for the front half while the reasoning arrives on a second meter. Fish Audio Live 3 does the same thing structurally: the speech layer runs on Fish Audio's Agent Platform while the agent's reasoning can point at any OpenAI-compatible endpoint through the custom-LLM hook. The difference is that OpenAI has published the first half's price and Artificial Analysis has published the whole system's score, while Fish Audio has published neither, and labels the one number it does give — 186 milliseconds, the shortest interval in which Live 3 can decide to react — as vendor-reported.

That makes this less a contest between two models than a contest between two ways of splitting a voice agent in half, and the split matters more than the model. GPT-Live-1 scored 81.5 on the Artificial Analysis Speech to Speech Index when paired with GPT-6 Astra at medium effort and 80.1 with GPT-5.6 Sol at low effort, which means the choice of backend moved the result by seven times the margin between first and second place. Live 3's equivalent number does not exist yet. Buying either one is buying half a system and committing to choose the other half yourself.

What each vendor actually published

GPT-Live-1 is the better-documented product, and the documentation is specific about the seam. It shipped inside ChatGPT on 2026-07-08 as the replacement for Advanced Voice Mode and reached the developer API on 2026-09-10. The Artificial Analysis board carries two rows for it because the index scores a whole system: 81.5 with GPT-6 Astra at medium effort, first place, and 80.1 with GPT-5.6 Sol at low effort, third. The component figures underneath the Astra row are a 67.9% agentic score on τ-Voice, 90% on Big Bench Audio, 87.4% task success, an arena Elo of 1,048, and 1.34 seconds to first audio — measured against OpenAI's own 0.798 seconds on Full-Duplex-Bench, which counts a different event.

Live 3's published surface is one paragraph and a set of worked examples. Fish Audio announced it on 2026-11-10, describes it as full duplex and trained on recorded human conversation, states that it can decide to respond in as little as 186 milliseconds for short reactions, says it is English only at launch, and offers it through the Fish Agent Platform, Fish Audio's own API and on-premises deployment. There is no benchmark, no τ-Voice figure, no arena Elo, and no line for it on Fish Audio's rate card — which lists drama-3, s2.1-pro, s2.1-pro-free, s2-pro and s1, plus three transcription models and voice-design-1.

• Front half — GPT-Live-1: $0.05 per voice minute, billed by the second, with 15 seconds charged at session start and credited against later usage. Live 3: price not published.

• Back half — GPT-Live-1: an asynchronous backend model of your choosing, billed at that model's rates, with GPT-5.6 Terra in OpenAI's published WebRTC example. Live 3: a custom-LLM hook pointing at any OpenAI-compatible endpoint, likewise billed by whoever serves it.

• Independent score — GPT-Live-1: 81.5 and 80.1 on the composite index, two configurations scored separately. Live 3: none, as of 2026-11-11.

• Transport and limits — GPT-Live-1: WebRTC only, no image or video input, no structured outputs, no free tier, and concurrency limits of 25, 50, 200, 300 and 500 simultaneous sessions across Tiers 1 to 5. Live 3: REST under /v1/agent/* with JavaScript and React SDKs, session-token auth, channels for a phone number, a web widget or a custom client.

Why the two-meter structure is the actual decision

A voice agent's bill has two halves and only one of them is predictable. The front half is metered by time: at $0.05 a minute, a five-minute call costs a quarter and an hour of open line costs three dollars, which is arithmetic simple enough to forecast. The back half is metered by tokens, varies with how hard the caller thinks, and is the half that produces the unpleasant surprises — which is why Artificial Analysis reports an all-in cost per hour rather than a per-minute voice rate. Its figure is $5.83 an hour for the Astra pairing and $4.47 for the Sol one, both including backend work, against the $3.00 an hour that the voice meter alone would suggest.

Live 3 has the same two-meter shape with one meter hidden. Fish Audio has published no rate for the speech layer, so the front half is unknown, and the back half is whatever the endpoint behind the custom-LLM hook charges. A ten-minute Live 3 call therefore has a computable cost only if you know both numbers, and today you know neither from the vendor.

That asymmetry is the reason a routing layer turns up in this comparison at all, and it belongs on the back half only. OrcaRouter does not host Fish Audio Live 3 or GPT-Live-1 — neither has a third-party voice endpoint, and the speech layer in either stack stays with its vendor. What OrcaRouter hosts is the half both systems delegate to: 216 models behind one key at provider list price with zero markup, automatic failover when a backend errors or times out, a routing DSL that decides which model handles which kind of turn, and model fusion so a cheap model can draft a tool call while a stronger one checks it before it fires. For a GPT-Live-1 deployment the practical effect is that the 1.4-point spread between the two scored configurations becomes something you can measure on your own traffic instead of read off a leaderboard. For a Live 3 deployment it means the reasoning layer is chosen before the speech layer's price is even known — which is the right order, because it is the half you can change later.

The latency number that keeps being compared to the wrong thing

Three latency figures circulate in this matchup and no two of them measure the same event. GPT-Live-1's 0.798 seconds comes from Full-Duplex-Bench, a conversational-dynamics suite that times a model inside a turn-taking scenario. Its 1.34 seconds is time to first audio, measured on the API. Live 3's 186 milliseconds is Fish Audio's own figure for the smallest interval in which the model can choose to emit a short reaction such as "mm-hmm" or "okay, 415" between the chunks of a phone number.

Read those together and it is tempting to conclude that Live 3 responds in a seventh of the time. That conclusion is not supported by anything either vendor published. The only comparison the figures allow is directional: both companies are claiming that the first audible event in a conversation can arrive inside a quarter of a second under some conditions, and neither has published a distribution. What GPT-Live-1 has that Live 3 does not is a third-party harness applying the same measurement to it and to its competitors, which is what makes 1.34 seconds a number a reviewer can check and 186 milliseconds a number a pilot has to verify.

What the architecture commits you to

The two halves are not equally portable, and that is the part worth thinking about before either pilot starts.

On the GPT-Live-1 side the voice layer is effectively fixed. Sessions run over WebRTC rather than the WebSocket transport most existing Realtime integrations use, the Live Sessions endpoint is OpenAI's alone, and probing the older path with a GPT-Live model ID returns an error saying sessions require WebRTC. Moving onto it is a transport rewrite. Moving off it is another one.

On the Live 3 side the speech layer is fixed in a different way: it is a beta platform from a smaller vendor, English only, with no published price and no independent score. What is portable there is the reasoning hook, and what is not is everything below it — the audio path, the phone number, the session lifecycle.

In both cases the mobile half is the reasoning model, and that is where a switch is a model-ID change rather than an integration project. The design lesson is the same on both sides of this comparison: put the intelligence in the layer you can swap, and keep the layer you cannot swap as thin as the product allows.

Which half to buy first

Buy the GPT-Live-1 half first if you need to justify the decision now. It has a published rate, two scored configurations, a component breakdown, a documented transport and a concurrency ladder you can size a call centre against. What it does not have is a video path, structured tool outputs, a free tier, or any independent reproduction of OpenAI's own 80.1% interactivity claim.

Pilot the Live 3 half first if your calls are English, your complaint log is full of callers being talked over, and on-premises deployment is a requirement rather than a preference. Go in knowing that you are the evaluation: no rate card line, no τ-Voice figure, no arena score, and a 186-millisecond claim that describes short reactions rather than answers.

Do not buy either half expecting the vendor to finish the product. GPT-Live-1 without a backend is a very good listener with nothing to say; Live 3 without a reasoning endpoint behind its hook is a beta speech layer with an unpublished bill. The score that decides this matchup — 81.5 against nothing — is really a statement about documentation, and the number that will settle it is the first independent turn-taking run on Live 3.