Fish Audio Live 3 Speaks One Language. Gemini 3.8 Live Speaks 97 and Watermarks Every Second.

Fish Audio Live 3 and Gemini 3.8 Live agree on the important idea and disagree about almost everything else. Both put turn-taking inside the speech model rather than in a pipeline of separate parts. Live 3, announced on 2026-11-10, argues the point behaviourally: a pause in the middle of a phone number is not the same event as the pause at the end of a question, and the model can decide to respond in as little as 186 milliseconds, a figure Fish Audio reports and nobody has audited. Gemini 3.8 Live, announced on 2026-09-15, answers the same problem with paperwork — 97 languages with mid-conversation switching, a SynthID watermark on every second of generated audio, a published model card, two documented API model IDs, and an Artificial Analysis row that reads 76.0 on the Speech to Speech Index with all four of its components filled in.

That asymmetry, not the latency figure, is what decides this comparison. One of these models can be written into a procurement document today and defended line by line; the other has to be instrumented by the team that adopts it. And there is a sharper point underneath: the behaviour Live 3 is selling is not the dimension Gemini 3.8 Live is strongest on, and it is not a dimension Live 3 has been scored on either — so for now the matchup is settled by what you can verify rather than by which model handles a phone call better.

The coverage gap is not a footnote

Start with what each model will do outside a demo, because for most buyers this section ends the evaluation before the interesting questions get asked.

• Languages — Gemini 3.8 Live supports 97 languages and switches between them automatically mid-conversation. Fish Audio states plainly that Live 3 is English only at launch. A booking line in Madrid, a support queue in São Paulo and a collections desk in Warsaw are not Live 3 deployments this quarter.

• Provenance — every second of audio Gemini 3.8 Live generates carries Google's SynthID watermark, and both Gemini voice variants are documented in a model card published under the name gemini-3-8-audio. Fish Audio's announcement says nothing about watermarking or provenance metadata for Live 3. If your jurisdiction is moving toward synthetic-voice disclosure, that difference arrives as a legal question rather than an engineering one.

• Identifiers — Google publishes gemini-3.8-live and gemini-3.8-live-extended-thinking as callable model IDs through the Gemini API, the Live API and AI Studio. Fish Audio's announcement names Live 3 but publishes no model identifier and no per-model endpoint documentation for it.

• Non-audio input — Gemini 3.8 Live processes visual input near real-time, on top of camera and screen-sharing support the line already had. Live 3 is audio in, audio out.

• Enterprise route — Google's path is private preview in Gemini Enterprise, with the customer-experience edition listed as coming soon. Fish Audio's is the Agent Platform, currently in public beta, plus on-premises deployment for enterprise customers — which is a genuinely different offer, and the one place in this list where the smaller vendor has the stronger answer for a buyer who cannot send audio to a public cloud.

What the index can see, and what it cannot

Artificial Analysis builds its Speech to Speech Index as an equal-weighted composite of speech reasoning, agentic performance, arena preference and task success, and scores a whole system rather than a single model. Gemini 3.8 Live's row is complete: index 76.0, Big Bench Audio 92%, time to first audio 1.18 seconds, task success 93.2%, conversational dynamics 96.1%, arena Elo 1,083, and 30.1% on the τ-Voice customer-service task benchmark. Live 3 has no row at all as of 2026-11-11; the board lists Fish Audio as a model family and scores none of its duplex work.

Look closely at two of those components, because they say something the headline does not. Gemini 3.8 Live's 96.1% on conversational dynamics is the highest figure in its row and the closest published proxy for the thing Live 3 is selling — a turn-based model scoring near the top of a turn-taking measure. Its 30.1% on τ-Voice is the lowest figure in the row and the reason the model sits fifth rather than first: it converses well and finishes tasks rarely. Google's own fix for that was a second variant, Gemini 3.8 Live Extended Thinking, which reaches 68.6% on the same test and takes first place on the index at 82.6 — for roughly four times the hourly audio cost.

The lesson for this comparison is uncomfortable for both vendors. Conversational dynamics and natural turn-taking are not the same measurement, which is exactly why a 96.1% on one and an unscored claim on the other cannot be lined up. Live 3's argument is that its behaviour is better than a turn-based model's in a specific situation — a caller pausing mid-number, or murmuring agreement while the agent reads an order back. Until someone runs it on a turn-taking harness, the honest position is that Live 3 is making a claim about a scenario the published boards do not isolate.

Provenance, and the question a reviewer will ask

Voice is the one modality where a compliance team asks about the output before it asks about the price. Gemini 3.8 Live has an answer ready: SynthID watermarking on generated audio, a named model card, and a vendor with a published responsible-use posture behind both. Live 3 has an announcement that is unusually precise about conversational behaviour and completely silent on how a listener or a detector could tell its output from a recording.

That is not a reason to avoid Live 3, and it is a reason to ask Fish Audio a direct question before you deploy it on a line that records calls. Whether the audio is watermarked, whether sessions produce a transcript or an audit record, and what the retention policy is on both are the three items a reviewer will want in writing. Fish Audio has documented the platform's session-token authentication and its /v1/agent/* REST surface, so the plumbing exists to answer all three — the company has simply not published the answers, and one day after launch that is the normal state of affairs rather than a red flag.

The price that exists and the one that does not

Gemini 3.8 Live is published at $0.005 per minute of audio input and $0.018 per minute of audio output through the Live API. A five-minute call with continuous output is therefore a little over eleven cents of voice, before any backend model cost. Artificial Analysis converts Google's rates into an hourly figure against a fixed forty-question subset, and that derived column has moved between captures — $0.84 an hour for the standard model in the 16 September board, $1.50 in a later one. Both are conversions, not quotes; the per-minute rates are the number to budget against.

Live 3 has no price. Fish Audio's rate card lists drama-3 at $50 per million UTF-8 bytes, s2.1-pro and s2-pro at $15, s2.1-pro-free at zero, and s1 at $15 with a retirement date of 2026-12-31, plus three speech-to-text models and voice-design-1 at a cent per request. Live 3 is not on it. A duplex agent call also has no natural byte count — the audio in is whatever the caller says — so even the shape of the meter is unpublished. Any Live 3 cost model written today is a guess, and the guess should be labelled as one in whatever document it lands in.

What sits behind the voice

Neither model is the whole system, and the seam is the same on both sides. Gemini 3.8 Live runs background tool and API execution, acknowledging a request and continuing to talk while the call completes — and that work resolves into ordinary model calls with ordinary token costs. Live 3's Agent Platform exposes a custom-LLM hook that points an agent's reasoning at any OpenAI-compatible endpoint. In both cases the speech model is a front end and a text model does the thinking.

That layer is where a router earns its place, and it is the only layer in either stack where switching is cheap. OrcaRouter carries 216 models behind a single key at provider list price with zero markup, so an upstream price cut lands the same day rather than at the next contract renewal, automatic failover keeps a call alive when a backend errors or times out, a routing DSL decides which model handles which kind of turn, and model fusion lets a cheap model draft a tool call while a stronger one checks it before it fires. To be precise about the boundary: OrcaRouter does not host Fish Audio Live 3 or the Gemini 3.8 Live voice endpoint. Neither has a third-party voice endpoint, and the speech layer in either stack stays with its vendor. The reasoning layer is the part that is portable, and for a stack built on a preview-priced model from one vendor and a beta platform from another, portability is the property that survives the next release.

What each one is actually for

Choose Gemini 3.8 Live if your product is global, if the audio leaves your infrastructure and someone will ask you to prove where it came from, or if you need the agent to see a screen or a camera. It has the scorecard, the watermark, the languages, the published rate and the model ID. Its weakness is task completion at 30.1% on τ-Voice, and the fix Google sells for that is the Extended Thinking variant at four times the hourly cost.

Pilot Fish Audio Live 3 if your calls are in English, your infrastructure constraints rule out a public cloud, and your complaint log is full of callers being talked over or repeating digit sequences. On-premises deployment is a real advantage here and one Google is not offering. Go in expecting to build the evaluation yourself: there is no score to compare against, no published price to forecast, and no watermark to point at.

Do not treat 76.0 as a verdict on turn-taking. It is a composite that weights agentic task completion heavily, and Gemini 3.8 Live is weak exactly there. The number that would settle this matchup is a turn-taking score for Live 3 on the same harness everyone else is measured on — and until that exists, the two models are being compared on documentation rather than behaviour.