Fish Audio Live 3 vs Grok Voice 3.0: One Price Is a Bundle and One Is a Component

Put the two rate cards side by side and the newer model looks less than half the price. Fish Audio Live 3 sells for $0.035 per minute on Fish Audio's Agent Platform — a promotional rate, half of a $0.07 list, good through 31 December 2026. Grok Voice 3.0 sells for $0.08 per minute, or $4.80 an hour, plus $0.004 per text input token. Four cents against eight is a clean story until you notice that the two numbers are billing different objects: Grok Voice 3.0 is a speech-to-speech component you rent, and Live 3's per-minute figure bundles speech recognition, the reasoning model and the voice into one meter. One is a part, the other is an assembly. Compare them per minute and you are comparing a gearbox to a car.

That distinction is the whole comparison, and it is worth being precise about why. Grok Voice 3.0 is a versioned model string behind a realtime WebSocket with session-scoped tools and returned transcripts, priced by the minute of audio it processes and scored, independently, at 86.9 on the Artificial Analysis Speech to Speech Index. Fish Audio Live 3 is a full-duplex model launched on 10 November 2026 that keeps listening while it speaks, is English-only at launch by the company's own statement, has no published benchmark of any kind, no model identifier in Fish Audio's API documentation, and no weight release — the vLLM project's public reply to the launch was to hope it gets open-sourced one day.

The scoreboard that exists on one side only

Artificial Analysis measures Grok Voice 3.0 on the same harness it uses for every duplex model, and the numbers are unusually complete. Every figure below is the board's own, taken from its run dated 13 October 2026.

• Speech reasoning, Big Bench Audio — 98.5%.

• Conversational dynamics, Full Duplex Bench — 97.0%, with backchannelling 92.9%, interruption handling 96.5%, turn-taking 100% and pause handling 98.5%.

• Speech Agent Arena preference — Elo 1059, against 1074 for Gemini 3.8 Live and 990 for Eleven Realtime v1 Max.

• Task success — 94.8% across the τ-Voice airline, retail and telecom scenarios.

• Time to first audio — 0.81 seconds, slower than Grok Voice 2.0's 0.70 and roughly 40% quicker than GPT-Live-1 in its Astra medium configuration at 1.34 seconds.

• Cost per hour of input audio — $4.80, a straight conversion of the vendor's own per-minute rate.

Live 3's column is not sparse, it is empty, and the emptiness has a specific shape. There is no row on that board for any Fish Audio duplex model as of 11 November 2026 — the board lists Fish Audio as a model family and scores none of its live work. Fish Audio's one published number is 186 milliseconds to decide to respond, which the company describes in the context of short reactions such as "mm-hmm" rather than full answers. It is a vendor floor on an acknowledgement. Grok's 0.81 seconds is an audited time-to-first-audio on a reply. Nothing about those two figures is commensurable, and the fact that both are expressed in seconds is the trap.

What each side's duplex claim actually commits to

Both models are full duplex, so "listens while it speaks" is not a differentiator between them; it is the entry ticket. The difference is where the timing policy comes from.

Fish Audio's argument is that acknowledgement, waiting, answering and stopping were learned from recorded human conversation rather than written as a list of silence thresholds, and it illustrates that with a phone number read out in chunks: an agent that hears three end-of-turn events where a human hears one hesitation is the failure being fixed. The supporting material is three sample calls, finishing an order, cutting in, and backchannelling through a read-back. What Fish Audio has not published is any measurement of whether the learned policy beats a thresholded one.

Grok Voice 3.0's timing policy is not explained in behavioural terms at all. What xAI published is a rate card, a model string, an alias move — grok-voice-latest now resolves to grok-voice-3.0, and grok-voice-2.0 is flagged deprecated — and, through Artificial Analysis, a set of turn-taking scores that a buyer can read as a risk statement. A 96.5% interruption score is a claim you can test against your own call recordings. A 186-millisecond acknowledgement floor is a claim you have to take on trust until you run it.

The alias move matters more than it sounds in a phone system. Grok Voice 2.0 became the alias target on 5 August 2026 and was superseded by 3.0 ten weeks later, which means a team that deployed against the alias has already been through two silent model swaps in a quarter. Whether you pin grok-voice-3.0 explicitly is a deployment decision the vendor has effectively forced.

Where the money actually lands

The bundled-versus-component split produces four cost shapes, and only one of them favours Live 3 in a way that survives contact with a real call.

• Live 3 on the Agent Platform — $0.035 per minute all-in during the promotion, covering recognition, reasoning and voice; add $0.01 per minute for phone calls, $1.20 a month per phone number, and the premium-tier LLM's token rate if you upgrade off the included model. A ten-minute call is 35 cents plus telephony.

• Live 3 after 31 December 2026 — the list rate is $0.07 per minute, at which point the all-in figure is within a cent of Grok's $0.08 and the bundle argument has to carry the comparison on its own. Fish Audio has not said what happens to the discount.

• Grok Voice 3.0 — $0.08 per minute of audio plus $0.004 per text input token, with the reasoning model, the telephony and the orchestration all bought separately. The per-minute rate is a floor, not a total.

• Either, with a reasoning model of your own — the only shape where the two converge, because both vendors let you supply the brain and bill the speech layer.

That last point is the practical conclusion hiding inside the price table. If your product differentiates on the reasoning layer, the bundle is a discount on something you were going to replace. If it does not, the bundle is genuinely cheaper and simpler, and the 50% promotion is the best deal in voice this month.

Procurement, and the question nobody's rate card answers

Price and scores are the easy half. The hard half is what happens when the model changes underneath you, and here the two vendors have made opposite trades.

Grok Voice 3.0 has a stable interface and an unstable model: same endpoint, same event names, same bearer-token handshake, same voice set, with the model swapped behind an alias. Fish Audio Live 3 has an unstable interface and, for now, a stable model: the platform is in public beta, the duplex model is new, its API documentation does not yet name it, and the launch post is the only document that describes its behaviour. One arrangement asks you to test continuously. The other asks you to read release notes.

Neither is unreasonable. But an engineering team choosing between them should notice which risk it is better equipped to absorb, because that answer is usually more predictive of a successful deployment than four cents a minute.

The portable layer, and the honest boundary

Both stacks are two-part systems, and in both the speech half is unportable — Grok Voice 3.0 runs only on xAI's realtime endpoint, Live 3 only on Fish Audio's platform and on enterprise hardware. The reasoning half is where a router belongs, and on both it is a separate purchase: on the Grok side it is a model you call from inside the session's tools, on the Fish side it is an OpenAI-compatible endpoint the agent points at. OrcaRouter carries more than 200 models behind one key at provider list price with zero markup, so the text model in either stack can be swapped or pinned without renegotiating the speech contract, and automatic failover keeps a session alive when the reasoning backend errors mid-call. Neither speech model is routable here — there is no third-party endpoint for either — and this piece would be misleading if it implied otherwise.

For a team still deciding, the cheapest experiment is the one the pricing makes possible: run the same twenty calls through both, score the interruptions yourself, and see whether 94.8% task success on someone else's retail scenarios predicts anything about yours. Live 3's 186 milliseconds will feel either like the point of the model or like a number about a noise the agent makes while it thinks. That is not a question a rate card can settle, and it is the only question that matters after the invoice arrives.