Grok Voice 3.0 vs Qwen-Audio-3.0-TTS: One of These Can Hear You, and It Is Not the Cheaper One

Set the two rate cards side by side and Alibaba's model wins by a factor of three. Qwen-Audio-3.0-TTS renders speech at $27.6 per million characters, roughly $1.66 of audio per hour, and it is available today as a hosted API through Alibaba Cloud Model Studio. Grok Voice 3.0, announced by SpaceXAI on 14 October 2026, bills $0.08 per minute of audio — $4.80 an hour — plus $0.004 per text input. The narrator is the cheaper product, and the narrator cannot hear anything you say. Qwen-Audio-3.0-TTS is a text-to-speech model. Grok Voice 3.0 is a full-duplex speech-to-speech model that has to decide, several times a second, whether you have finished talking. Per hour of audio is the right metric for exactly one of them.

That is the whole matchup in one paragraph, and everything below it is the part that decides deployments: what each one costs to operate once you count the legs of the pipeline around it.

One renders a script. One takes a call.

Text-to-speech takes text and returns audio. There is no input audio, no turn to take, no concept of being interrupted and no transcript coming back, because the model never heard anything. Its cost model is a printing press — pages in, audio out — and the only two numbers that matter are quality and throughput, both of which you can measure before you ship.

A full-duplex model processes incoming audio while it is producing outgoing audio. Its job includes the parts of a conversation that have nothing to do with words: pausing when the caller pauses, continuing through a backchannel rather than treating it as an interruption, yielding the floor mid-sentence, and firing a tool call without dropping the thread. Grok Voice 3.0 does all of that and returns transcripts alongside the session, which it can only do because it was listening the whole time.

If your product reads articles aloud, converts documentation to audio, or narrates video, Grok Voice 3.0 is the wrong tool at three times the per-hour price and cannot do the job at all, because it does not accept a script and produce a file. If your product takes a phone call, Qwen-Audio-3.0-TTS is one third of a pipeline that still needs speech recognition and a language model to become a conversation — and each of those legs adds cost, latency and a boundary where the prosody dies.

Where Qwen-Audio-3.0-TTS is genuinely strong

The case for Alibaba's model is not marketing. Released on 20 July 2026 by Alibaba's Tongyi Lab and exposed as a hosted, closed API, the Plus tier took the top spot on Artificial Analysis's text-to-speech arena when it arrived. Leaderboards move, and this one has: as of September 2026 Qwen-Audio-3.0-TTS-Plus sits fourth on the Provider Voice Arena with an Elo of 1,232 over 2,306 samples, behind Cartesia Sonic 3.6 at 1,275, Inworld Realtime TTS-2 at 1,244 and Speechify's Simba 3.2 at 1,233 — three models from July and August that shipped after it. Fourth of twenty-plus, with the lead gone and the price advantage intact, is still a strong position. It is just not the position the launch coverage described.

• Language coverage — 16 languages with lead positions in 10 of them, plus 20 Chinese dialect regions.

• Voice cloning — from short reference samples, including cross-lingual cloning.

• Expressive control — 86 inline emotion and delivery tags.

• Throughput — Plus generates roughly 16 characters per second. For batch rendering that is invisible; for a live product it is the number that decides your buffer.

• Tiers — Flash targets roughly 300 ms first-packet latency for real-time playback; Plus is the quality tier. They are different products sharing a name, and picking the wrong one is a common first mistake.

• Documentation gap — the widely cited $27.6 per million characters does not match Alibaba's own pricing page in every snapshot, which has at points listed lower figures for both tiers. Check the number at account level before you forecast rather than the leaderboard's.

Two caveats worth planning around. Despite sixteen supported languages, the public preset voices are almost entirely Chinese and English, and the other fourteen exist through the cloning workflow rather than off the shelf. And the 86 inline emotion tags work in unidirectional streaming mode only, so the expressive control does not survive every integration path.

There is also a version to be careful about. Alibaba now lists a Qwen-Audio-3.1-TTS generation and a separate Qwen Audio 3.0 Realtime family, and those are not the same model as Qwen-Audio-3.0-TTS — the Realtime line is a speech-to-speech product that appears on a different leaderboard entirely. Anyone comparing "Qwen Audio 3.0" against a conversational model should check which row of Alibaba's own price list they are actually reading.

What Grok Voice 3.0 costs, and what it does not solve

Grok Voice 3.0 landed on 14 October 2026 at $0.08 per minute of audio plus $0.004 per text input, unchanged from the model it replaces, with grok-voice-latest now resolving to it and grok-voice-2.0 flagged deprecated. On Artificial Analysis's Speech to Speech Index it holds the top row at 86.9, its highest component being agentic performance at 71.6% on τ-Voice against 56.5% for its predecessor — the largest single move in the release.

The latency claim is where the two models finally share a unit. Grok Voice 3.0 publishes 0.81 seconds to first audio, the fastest figure among frontier conversational models. Qwen Audio's Flash tier targets roughly 300 ms first packet. Those are not the same measurement: 300 ms is time to the first packet of a rendering job whose text you already have, while 0.81 seconds is the time to start speaking after a human stopped talking, which includes the decision that it was your turn. Roughly two and a half times the latency buys the ability to be interrupted. Whether that is worth it depends entirely on whether a person is on the other end.

Grok Voice 3.0 also brings the parts of a voice product that a TTS model leaves to you: the duplex session, server-side turn detection, session-scoped tools so the agent can call a backend mid-conversation, and the transcription that comes back with it. It does not bring a script-reading mode, cloning that is documented as cross-lingual, or sixteen languages marketed as a feature. Its addressable use case is narrower than the TTS model's, and its per-hour number is higher.

The honest version of the cascade bill

A conversational product built on a text-to-speech model is a cascade: a speech-recognition model turns the caller's speech into text, a language model decides what to say, and the TTS model speaks it. Three vendors, three bills, three latency budgets that add up, and a hand-off at each boundary. The TTS leg may cost $1.66 an hour of audio. The other two legs are where the money and the mistakes actually are, and the cascade cannot hear a pause in the middle of a sentence it is already speaking.

Grok Voice 3.0 collapses the three legs into one session and one meter — with two details that keep the bill honest. Its tool calls and any delegated reasoning run on whatever backend answers them, at that model's own rates, so pointing the agent at a frontier reasoner that searches on every turn produces an expensive call behind a cheap voice layer. And at 300,000 minutes a month, the rate card is $24,000 on the voice half alone.

That second meter is where this comparison stops being a model choice and becomes an architecture choice, and it is where a routing layer earns its place. OrcaRouter carries neither Grok Voice 3.0 nor Qwen-Audio-3.0-TTS; the voice layers in both cases are outside our reach and we route none of it. The reasoning leg is not. Two hundred and seven models from a dozen-plus upstreams sit behind one key at provider list price with zero markup, and a provider price cut lands the same day rather than at your next contract renewal. For a cascade that covers the middle leg outright: hold two speech-recognition options and three reasoners behind one endpoint, A/B them against real traffic, and let automatic failover absorb an upstream outage without dropping a call. A routing DSL composes several models into one call for the cases where a single reasoner is not enough.

Choosing, in one question

Buy Qwen-Audio-3.0-TTS if the output is content: narration, localisation, accessibility audio, dubbing, or any workflow where the text exists before the audio does and quality per rendered hour is the metric. Start on Flash if you need live playback and move to Plus when the voice itself is the deliverable, price the cloning workflow separately from the preset voices, and confirm the per-million-character rate at account level before you build a forecast on it.

Buy Grok Voice 3.0 if the input is a person. Support lines, booking flows, intake interviews, anything where the caller's first syllable arrives while the model is mid-sentence and the system has to behave like a listener rather than a player. It costs more per hour of audio, it cannot read your document, and it is the only one of the two that can be interrupted.

Plenty of real products want both — a rendering model producing the audio assets and a duplex model handling the call those assets belong to. What they should not share is a cost model. Per hour of audio is the right unit for one of these two, and the other one's real cost is the pipeline you avoided by not building a cascade.