GPT-Live-1 vs Qwen-Audio-3.0-TTS: A Conversation and a Narrator Are Not the Same Line Item

Put GPT-Live-1 and Qwen-Audio-3.0-TTS side by side on price per hour of audio and the gap looks decisive: Aliba​ba's Qwen-Audio-3.0-TTS-Plus renders speech at $27.6 per million characters, roughly $1.66 of audio per hour, while Ope​nAI's GPT-Live-1 bills $3.00 for an hour of continuous open line. The narrator is cheaper, so the narrator wins — except that this is the wrong comparison, and the reason it is wrong is the most useful thing anyone can take from the matchup. Qwen-Audio-3.0-TTS is a text-to-speech model. GPT-Live-1 is a full-duplex conversational model. One of them reads what you wrote. The other one has to decide, several times a second, whether you have finished talking.

One renders, one converses

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

A full-duplex conversational 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 user pauses, continuing through a backchannel like "yeah" instead of treating it as an interruption, yielding the floor mid-sentence, and firing a tool call without dropping the thread. GPT-Live-1 does all of that and returns speech-recognition 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 a video, GPT-Live-1 is the wrong tool at four times the per-hour price. If your product takes a phone call, Qwen-Audio-3.0-TTS is one third of a pipeline that still needs an ASR model and a language model to become a conversation.

Where Qwen-Audio-3.0-TTS is genuinely the best answer

The case for Aliba​ba's model is not marketing. Released on July 20, 2026 by Aliba​ba's Tongyi Lab and exposed as a hosted, closed API through Aliba​ba Cloud Model Studio, the Plus tier took the top spot on Artificial Analysis's text-to-speech arena when it arrived. A leaderboard is a moving target, though, and this one has moved: 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 August and July 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.

It leads in 10 of the 16 languages it covers, adds 20 Chinese dialect regions, supports voice cloning from short samples including cross-lingual cloning, and carries 86 inline emotion and delivery tags.

The caveats are specific enough to plan around.

• Throughput — Plus generates roughly 16 characters per second, against 30.2 for Simba 3.2, 27 for Gemini 3.1 Flash TTS and around 120 for Sonic 3.5. For batch rendering this is invisible; for a live product it is the number that decides your buffer.

• Price documentation — the widely cited $27.6 per million characters does not match Aliba​ba's own pricing page in every snapshot, which has at points listed lower figures for both tiers. Check the current number at the account level before you forecast, not the leaderboard's.

• Voice library — despite 16 supported languages, the public preset voices are almost entirely Chinese and English. The other fourteen languages exist through the cloning workflow rather than off the shelf.

• Feature gating — the 86 inline emotion tags work in unidirectional streaming mode only, so the expressive control does not survive every integration path.

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

The honest version of the cascade bill

Here is what the per-hour comparison leaves out. A conversational product built on a TTS model is a cascade: an ASR 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 where prosody dies. 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 model cannot hear a pause in the middle of a sentence it is already speaking.

GPT-Live-1 collapses the three legs into one session and one meter, at $0.05 per minute of voice with the reasoning backend billed separately. Two details keep that bill honest. Session initialisation bills 15 seconds of voice up front, credited against later duration rather than added to it. And the backend is a second meter: every delegated reasoning call, tool invocation and web search bills at that model's normal rates, so pointing delegation at a frontier reasoner that searches on every turn produces an expensive call behind a cheap voice layer.

What the independent boards say about the two is instructive precisely because they measure different things and neither flatters its subject. Qwen-Audio-3.0-TTS-Plus is fourth on quality and near the bottom on throughput. GPT-Live-1 is seventh of eleven on Artificial Analysis's Speech to Speech Index at 69.8% — behind the GPT-Realtime-2.1 it replaces, at 73.9% — while being charged at the lowest end of the index's cost-per-hour-of-input-audio range, $4.42 against GPT-Realtime-2.1's $10.75. Neither model wins its own category's top spot. Both are cheaper than the thing they were built to beat.

That second meter is where a routing layer becomes a design decision rather than an optimisation. OrcaRouter carries neither GPT-Live-1 nor Qwen-Audio-3.0-TTS — the voice layer in both cases is outside our reach, and we route none of it. The reasoning leg is not. Roughly 190 models from eleven upstream providers sit behind one key at provider list price with zero markup, and a provider price cut lands the same day rather than at the next contract renewal. For a cascade, that covers the middle leg outright: hold two ASR options and three reasoners behind one endpoint, A/B them against each other on real traffic, and let automatic failover absorb an upstream bad afternoon without dropping a call.

The consent question cuts the other way

Voice cloning is Qwen-Audio-3.0-TTS's most commercially useful capability and its largest compliance surface. Ten seconds of reference audio is enough to reproduce a speaker, cross-lingually, which is transformative for localisation and a liability anywhere identity matters. Ope​nAI shipped GPT-Live-1 with no cloning and no custom voices at all — 12 API voices to choose from, and that is the list.

That asymmetry is easy to skip past in a feature comparison and hard to skip past in a risk review. A product that only ever speaks in one of twelve vendor voices has a much shorter answer for "whose voice is that, and who consented to it" than a product that can reproduce any voice from a sample. If you need cloning, the pipeline decision is already made for you, and the question becomes what governs it. If you do not, GPT-Live-1's limitation is worth reading as a control you did not have to build.

Which one to buy

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 real-time playback, move to Plus when the voice itself is the deliverable, and price the cloning workflow separately from the preset voices.

Buy GPT-Live-1 if the input is a person. Support lines, booking flows, intake interviews, anything where the user'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, and it is the only one of the two that can be interrupted.

The two are complements far more often than rivals: plenty of products want Qwen-Audio-3.0-TTS reading a document and a duplex model handling the call that follows. What they should not share is a cost model. Per-hour-of-audio is the right metric for exactly one of them.