Fish Audio Live 3 vs Eleven Realtime v1 Max: The 186 ms and the 0.57 s Are Not the Same Measurement

Fish Audio Live 3 and Eleven Realtime v1 Max are both full-duplex voice models, both shipped within four weeks of each other, and both are sold on the claim that turn-taking belongs inside the speech model rather than in a pipeline of separate parts. Fish Audio says Live 3 can decide to respond in as little as 186 milliseconds. ElevenLabs' Realtime v1 Max starts talking in 0.57 seconds. Put side by side those two figures look like a three-fold win for the newer model, and that reading is wrong in a way worth understanding before anyone signs a contract.

The numbers measure different events. Live 3's 186 milliseconds is Fish Audio's own figure for short reactions — an "mm-hmm", an "okay, 415" between the chunks of a phone number — not for a full answer. Max's 0.57 seconds is a measured time-to-first-audio for the response itself, taken on Artificial Analysis' Speech to Speech harness. One is a vendor-reported floor on an acknowledgement; the other is an audited latency on a reply. Subtract them and you get a number that means nothing, which is the useful starting point for this comparison.

What each vendor has actually published

Eleven Realtime v1 Max arrived on 2026-10-20 with a full scorecard. Artificial Analysis scores it on Big Bench Audio (98.8%), time to first audio (0.57 s), task success rate (87.1%), Full-Duplex-Bench conversational dynamics (94.1%) and Arena preference (997 Elo). It does not have a Speech to Speech Index because the agentic τ-Voice test behind that index has not been run on it, and it does not get a cost-per-hour figure from ElevenLabs either — the $4.80 an hour on the board is a conversion of ElevenLabs' per-minute agent and API rates, not a published list price for the model alone.

Fish Audio Live 3 arrived on 2026-11-10 with a behaviour argument and almost no numbers. The published surface is small enough to state completely.

• Reaction speed — 186 milliseconds, vendor-reported, described in the context of short reactions rather than full answers.

• Independent scores — none. Artificial Analysis carries no Live 3 row on its Speech to Speech board as of 2026-11-11; Fish Audio appears there as a model family, not as a scored duplex entry.

• Price — not published. Fish Audio's rate card lists drama-3, s2.1-pro, s2.1-pro-free, s2-pro and s1, plus three speech-to-text models and voice-design-1. Live 3 is absent from it.

• Languages — English only at launch, per the announcement.

• Distribution — the Fish Agent Platform, Fish Audio's own API, and on-premises deployment for enterprise customers.

• Architecture — full duplex, trained on recorded human conversation so that acknowledging, waiting, answering and stopping are learned behaviours.

ElevenLabs' published surface is the mirror image: a complete scorecard and an opaque unit price. Max has every row except the index and the τ-Voice test, and ElevenLabs does not sell it as a standalone per-minute rate — in ElevenAgents you pay the standard agent charge plus tokens at the Realtime rate, and through the Realtime API you pay per minute of connected audio plus tokens.

The behaviour argument, which is the real competition

Both vendors describe the same failure mode and propose the same structural fix, which makes the comparison unusually clean at the level of ideas. ElevenLabs says recognition, turn-taking, reasoning and voice generation in Realtime are tuned as one system rather than coordinated across parts, and its documented example of the payoff is the backchannel: Mini is described as detecting whether a speaker has finished or is merely interjecting with an "mm-hmm". Fish Audio's version of the same claim is a phone number read in chunks — "415… 555… 0132" — where Live 3 might answer "mm-hmm" or "okay, 415" between them and wait for the rest.

Where they diverge is in what they will let you keep. Eleven Realtime is sold two ways, inside ElevenAgents and through a Realtime API that is wire-compatible with OpenAI's event protocol at wss://api.elevenlabs.io/v1/convai/realtime, and choosing a Realtime model inside ElevenAgents means ElevenLabs manages the LLM, reasoning effort, thinking budget, parallel tool calls, TTS model and ASR provider. Fish Audio's Agent Platform, currently in public beta, takes the opposite stance on the reasoning layer: an agent's custom LLM can point at an OpenAI-compatible endpoint, so the model doing the thinking is yours to choose.

That difference decides more deployments than the latency figures do. If you want the vendor to own the whole stack and you are willing to write off custom-LLM and Bedrock routes, ElevenAgents is the shorter path. If the reasoning layer is where your product differentiates — retrieval, policy checks, tool orchestration — the custom-LLM hook on the Fish side is the feature that survives procurement.

Cost, and why the Live 3 column is empty

Max costs $4.80 an hour of input audio on the Artificial Analysis conversion, against a $0.84 figure for the cheapest strong duplex competitor on that board. That makes price the axis where a new entrant can win a deal quickly, and it is exactly the axis on which Live 3 currently cannot be evaluated. A per-call cost for a Live 3 deployment cannot be computed from anything Fish Audio has published, because the billing unit is not published.

What Fish Audio does publish is a rate card for its other models, and it is worth knowing the shape of it because it tells you how the company prices speech. Text-to-speech is billed by the UTF-8 byte of input text, and Fish Audio's own documentation states that one million UTF-8 bytes is roughly 180,000 English words or about twelve hours of speech. At drama-3's list rate of $50 per million bytes that works out to roughly $4.17 per hour of generated speech, or about $0.83 during the launch promotion that runs through 2026-11-30. Those are derived figures for a different product, not Live 3 pricing, and duplex agent minutes are not the same unit as synthesized speech — but they put Fish Audio's price scale in the same $1-to-$5-per-hour range as the rest of this market rather than an order of magnitude below it.

This is the section where a routing layer earns its keep on both sides. OrcaRouter does not host Fish Audio Live 3 or Eleven Realtime v1 Max — neither has a third-party endpoint, and the speech layer in either stack stays with its vendor. What OrcaRouter hosts is the reasoning layer each of them can be pointed at: 216 models behind one key at provider list price with zero markup, automatic failover when a provider errors, and a routing DSL that decides which model handles which call. The practical consequence is that neither vendor's token price list becomes a lock-in. If ElevenLabs reprices Realtime, or Fish Audio publishes Live 3 as a per-byte rate, the speech bill changes and the reasoning bill does not, because the reasoning bill was never tied to either vendor to begin with.

Which one to build on now

Build on Eleven Realtime v1 Max today if you need to defend a technical choice to someone who will ask for a benchmark. It has an audited scorecard, a documented event protocol, and a measured latency that a reviewer can check. The unknown on Max is price per hour at volume, not quality.

Pilot Fish Audio Live 3 if your agent's complaint log is full of callers being talked over, and if you are prepared to measure the improvement yourself. The 186-millisecond figure cannot be compared to Max's 0.57 seconds, the model has no independent turn-taking score, and its list price does not exist yet — so a pilot has to be structured as an instrumented A/B against your current agent, not as a procurement decision on a spec sheet.

Do not treat 186 milliseconds as a latency claim you can hold anyone to. It is a vendor statement about how fast the model can choose to make a short noise. The number you actually care about — how long until the caller hears the first syllable of the answer — has not been published for Live 3 by anyone. When it is, the comparison with Max becomes a real one, and until then the honest answer to "which is faster" is that one of them is measured and the other is described.