Fish Audio Live 3 vs Drama 3: One Model Reads the Script, the Other Hears the Room

Fish Audio Live 3 and Fish Audio Drama 3 are not two answers to the same question, and treating them as rivals is the fastest way to buy the wrong one. Drama 3, which Fish Audio published on 2026-10-13, is a text-to-speech model: you hand it words and a performance direction, and it returns audio. Live 3, announced on 2026-11-10, is a duplex voice model for agent calls: it listens while it speaks, decides when a pause means "keep going" and when it means "I'm done," and can react in as little as 186 milliseconds according to Fish Audio. One is a renderer; the other is an interlocutor.

The confusion is understandable because both are Fish Audio models, both are recent, and both are sold on being good at something the other Fish Audio models were not. But they sit at different layers of a voice stack, they are documented separately, and only one of them appears on Fish Audio's published rate card. The useful comparison is therefore not which is better — it is which job each one is actually for, and where they meet inside one product.

What each model is documented to do

Drama 3 is a text-to-speech model, and the public evidence for it is preference and pronunciation rather than turn-taking. On Artificial Analysis' Provider Voice Arena it sits sixth at an Elo of 1,278; on the Controlled Voice Arena, where every model is judged against the same reference voices, it is twelfth at 1,085. On both boards it is the highest-ranked open-weights model in the table, and the earlier Drama 3 Preview record — the one that got the press in October — is still listed separately at 1,240 and 1,063. On the English Pronunciation Robustness benchmark it scores 93.2% overall, joint first among 31 scored models with Ming-3.0-tts, with the category breakdown mattering more than the headline: 95.9% on standalone terms such as brands and technical names, 93.6% on expanding shorthand such as numbers and dates, 86.3% on preserving exact sequences such as codes and paths, and 95.3% on contextually appropriate pronunciation.

Live 3's public evidence is a different shape entirely. Fish Audio's announcement is built around conversation rather than rendering, and its three worked examples are all about timing.

• A phone number read in chunks — "415… 555… 0132" — where Live 3 might say "mm-hmm" or "okay, 415" between the chunks and then wait for the rest.

• A caller saying "yeah" or "mm-hmm" while the agent reads out opening hours, where Live 3 keeps going instead of treating the acknowledgement as an interruption.

• A caller revising an order mid-utterance — "two pizzas… no, three. Three large. And… actually, make one of them a margherita" — where Live 3 waits through each correction instead of acting on the first plausible complete order.

Fish Audio states the model can decide to respond in as little as 186 milliseconds, describing short reactions such as "mm-hmm", "okay" and quick confirmations rather than full answers. It speaks English only at launch and is available on the Fish Agent Platform, through Fish Audio's API, and on-premises for enterprises. What it does not have is any published benchmark, any independent score — Artificial Analysis carries no Live 3 duplex row as of 2026-11-11 — or a line on Fish Audio's rate card.

Why the pricing units tell you they are different products

Fish Audio bills Drama 3 in UTF-8 bytes of input text: $50.00 per million bytes at list, with a launch promotion of $10.00 per million running through 2026-11-30. The company's own documentation converts one million UTF-8 bytes to roughly 180,000 English words, or about twelve hours of speech — which puts drama-3 at roughly $4.17 an hour of generated audio at list, and around $0.83 an hour during the promotion. That unit exists because TTS has a knowable amount of text going in and a knowable amount of audio coming out.

Live 3 does not fit that unit at all, and it is absent from the rate card. A duplex agent call has no fixed script: the audio in is whatever the caller says, the audio out is however much the model chooses to say, and the interesting cost is minutes of connection rather than bytes of text. Fish Audio has not published which meter it will use, which is why the Live 3 column in any cost comparison is blank rather than cheap.

What the two models do share is a platform. Fish Audio's Agent Platform, currently in public beta, hosts Live 3 agents and exposes a REST surface under /v1/agent/* with JavaScript and React SDKs, channels for a web widget, a phone number or a custom client, and session-token authentication. The custom-LLM hook on that platform points at an OpenAI-compatible endpoint, which is the seam where a reasoning layer of your own choosing goes.

Where they meet in one product

The interesting question is not which model wins but what a single product looks like when both are used for what they are good at. A booking line needs Live 3 on the live call, because the call is where pauses and corrections happen. The same product usually also needs Drama 3, or a model like it, for everything that is not a conversation: the confirmation read back after the call, the voicemail greeting, the outbound reminder, the long-form narration in the marketing page that sent the caller there. Those are render jobs with a known script, and Drama 3's published strengths — preference among listeners, 93.2% pronunciation, 83 languages, the open weights published under the Fish Audio Community License v1.0 with a $500,000 revenue cap — are exactly the strengths a render job needs.

That split has a cost consequence worth stating plainly, because it is where a routing layer stops being an abstraction. The reasoning layer behind a Live 3 agent can point anywhere that speaks the OpenAI protocol; OrcaRouter sits behind that endpoint with 216 models from a single key, at provider list price with zero markup and automatic failover when a provider errors, so the planner inside the agent can be swapped without touching the speech layer or the phone number. A routing DSL decides which model handles which kind of call, and model fusion lets a cheap model draft a tool call while a stronger one checks it before it fires. OrcaRouter does not host Fish Audio Live 3 or Drama 3 — neither is available through a third-party endpoint, and the speech layer in this stack stays with Fish Audio. What it hosts is everything around the speech.

Which one you actually need

You need Drama 3 if the output is a file. Narration, audiobooks, ads, character voice, dubbing, accessibility reads — anything where a human wrote the words and a listener will hear them once, polished. It has the preference scores, the pronunciation breakdown, the language count and a published per-byte rate you can forecast against. It is also open weights, which matters if you need to run it yourself.

You need Live 3 if the output is a conversation that has to survive being interrupted. Phone agents, booking lines, order taking, support triage — anything where the caller talks back and the failure mode is being talked over or having to repeat a digit sequence. It has a behaviour argument, a vendor latency floor and a platform, and it does not yet have a benchmark or a price.

You probably need both if you are building a voice product rather than a demo, and the split is along the line between live and rendered. The mistake to avoid is picking one to do the other's job: a TTS model cannot decide when to stop talking, and a duplex agent model is an expensive and unevaluated way to produce a voicemail greeting. Fish Audio sells both, the pricing units are different on purpose, and the roadmap that treats them as one product will mis-forecast its own costs from the first month.