GPT-Live-1 對決 Qwen-Audio-3.1：全雙工之爭才是真正值得打的一仗

The obvious way to write this comparison is price per hour of audio, and it is the wrong way. Q​wen-Audio-3.1-TTS renders speech at $19.30 per million characters on Artificial Analysis's conversion — roughly a fifth of what an hour of G​PT-Live-1 voice time costs at $0.05 a minute — and that number will end up in a procurement deck this month. It should not, because Q​wen-Audio-3.1 is not one model. It is a five-model refresh A​libaba announced on 23 September 2026 at the Apsara Conference in Hangzhou, and one of those five, Q​wen-Audio-3.1-Realtime, does the same job G​PT-Live-1 does. That is the matchup with an actual decision inside it, and A​libaba cut its price by about 85%.

兩個家庭，並排

A​libaba's refresh covered synthesis, recognition, realtime conversation and audio understanding, with cuts running across the whole line: synthesis down about 70%, realtime down about 85%, recognition as much as 95%. The tiers are not interchangeable and each has its own price and job.

• Q​wen-Audio-3.1-TTS — the synthesis successor, 16 languages including seven added in this release, 20 Chinese dialect regions retained, instruction-based control of emotion, pace and delivery, and natural cross-language timbre transfer that carries one voice across Mandarin, Cantonese, English and Japanese.

• Q​wen-Audio-3.1-TTS-Next — a different product wearing a familiar name. A​libaba calls it an audiogen model: voice, sound effects and background ambience generated in a single pass from text, timestamps and reference audio, with multi-speaker dialogue and 48 kHz output. Per A​libaba Cloud's Model Studio documentation it takes up to 3,000 characters of input, caps generated audio at 240 seconds for podcasts and 120 otherwise, runs at 3 requests per second, and returns WAV, MP3 or PCM.

• Q​wen-Audio-3.1-Realtime — full-duplex conversational, and the one A​libaba is pushing into hardware: Q​wen Office, Qoder, QwenNote A2, the QwenNote Eva desktop robot, and the Q​wen AI glasses.

• Q​wen-Audio-3.1-ASR and ASR-Next — recognition, and audio understanding covering emotion, music, environmental sound and event localisation.

G​PT-Live-1 is one thing, not a family. Full-duplex voice sessions, audio and text in, audio and text out, images and video unsupported, knowledge cut-off 31 July 2025, served only through v1/live/sessions with Chat Completions, Responses, Realtime and Batch all out of scope. It streams and supports function calling; structured outputs and fine-tuning are not supported. Its architecture is a split: G​PT-Live handles the conversation while a delegated backend model or agent handles reasoning and tools, either through Responses delegation with an O​penAI model or client delegation to your own service.

所以，誠實的對比是 TTS 對 TTS、Realtime 對 Live。這兩半的定價方式截然不同。

The rendering half: where Q​wen is genuinely hard to argue with

If your workload is narration, documentation audio, IVR prompts or video voice-over, Q​wen-Audio-3.1-TTS is a strong default and the price cut made it stronger.

• Independent standing — Elo 1,296 on Artificial Analysis's Provider Voice Arena, third of the field behind Eleven's two entries at 1,330 and 1,329, and ahead of Cartesia Sonic 3.6 at 1,285, Gemini 3.8 Flash TTS at 1,275, Q​wen's own previous generation at 1,265 and Realtime TTS-2 at 1,262.

• Controlled Voice 的排名 — Elo 1,186，在克隆語音排行榜上位居第二，落後於 HeyGen Voice 的 1,202，擁有 1,856 個樣本，而 HeyGen 為 1,482 個。

• 競技場的換算價格——每百萬字元 19.30 美元，相較於 Eleven v4 Turbo 的 40.00 美元、Gemini 3.8 Flash TTS 的 16.5 美元，以及 Cartesia Sonic 3.6 的 49.0 美元。

• 涵蓋範圍 — 16 種語言、20 個中文方言區、跨語言音色轉換，以及基於指令的情緒與表達控制。

• 語音庫 — TTS-Flash 上有 68 個語音，分為四種多語言與方言語音、22 種高級華語語音、15 種高級英語語音，以及 27 種額外系統語音。語音 ID 區分大小寫，且可用性並不統一：僅多語言與高級英語語音集由新加坡區域提供，其餘則僅限北京。

One numbering caution worth carrying into a budget: A​libaba publishes token rates while Artificial Analysis publishes a per-character conversion, and the two do not reconcile. A​libaba's own documentation lists TTS-Flash at roughly $0.23 per million tokens in and $1.87 out in Singapore, which is a different unit again from both the arena's $19.30 per million characters and the older $27.6 figure that still circulates for the previous generation. Check the rate at account level in the region you actually deploy to, not the one in a screenshot.

Against G​PT-Live-1, the rendering comparison resolves the same way it does against every TTS model: O​penAI's engine bills by open minute, whether or not anything is generated, and cannot be batched. For a scripted workload that is roughly twenty-seven times the per-audio-minute cost of the cheapest cartel-tier renderers and structurally the wrong tool.

對話的那一半，才是實際的決策

兩家廠商現在都販售全雙工對話模型，而各自背後的證據屬於不同種類的證據。這種不對稱比任何單一數字都更重要。

G​PT-Live-1 has an independent placement. Artificial Analysis's speech-to-speech index — a weighted average of speech reasoning, agentic performance, arena preference and task success — lists it under named backend configurations rather than as one row, with G​PT-Live-1 (Astra, medium) and G​PT-Live-1 (Sol, low) among the top entries, below G​rok Voice 3.0 and alongside G​emini 3.8 Live Extended Thinking. The cost panel on the same board prices G​PT-Live-1 configurations between roughly $4.42 and $10.75 per hour of input audio. That is a genuine third-party evaluation, and it is also a reminder that in this architecture the backend moves the score: the same voice model appears twice with different results.

Q​wen-Audio-3.1-Realtime does not appear on that index. Its standing rests on A​libaba's own benchmark claims for the family, its deployment into A​libaba's consumer hardware, and the price cut. Absence from a leaderboard is not a quality verdict — the board requires a specific set of component results before it publishes an index score — but it does mean that on the conversational side, one of these two has an independent number and the other has a vendor's.

The commercial difference is starker than the technical one. A​libaba cut realtime pricing by roughly 85% in the same announcement, which is the kind of move that changes what a product can afford to be. A Chinese-market consumer device shipping a full-duplex model at that rate is a different business than one shipping it at $0.05 a minute.

兩家供應商都不會替你解決的那件事：推理究竟發生在哪裡

Both systems delegate. G​PT-Live-1 does it explicitly and by design: it listens and speaks while a backend model or agent reasons, uses tools and returns results for it to communicate, and the backend is a separate bill at normal model rates on top of the $0.05 a minute. A​libaba's stack does the analogous thing across a five-model family, which is why the ASR, audiogen and realtime tiers exist as separate line items.

The consequence is the same in both cases and it is where most conversational projects overspend. Full session time is billable on G​PT-Live-1 including while the backend works. O​penAI's own cost guidance makes the arithmetic explicit — running a faster backend with parallel tool calls so the user finishes a minute earlier saves $0.05 in voice charges — and the lever is session length, not the voice rate. A conversation whose backend stalls for four seconds is a conversation that costs more and feels worse at the same time.

This is the part of the stack a routing layer actually changes. Client delegation means the delegated backend can be your own harness, so the reasoning model becomes a variable you can tune per route instead of a fixed dependency. OrcaRouter fronts 200-plus models at provider list price with no markup, which matters most exactly when a vendor cuts a price — the cut is live here the same day rather than at your next contract renewal. Automatic failover keeps a provider outage out of a live conversation, and the routing DSL lets you compose the reasoning step from several models rather than betting the session on one. To state the boundary plainly: OrcaRouter hosts neither G​PT-Live-1 nor any Q​wen-Audio-3.1 tier. What it routes is the delegated reasoning behind a Live session.

哪一個，用於哪項工作

Scripted audio at volume: Q​wen-Audio-3.1-TTS, and check the current tier pricing at account level rather than trusting the arena's conversion. It is third on the independent provider board, second on the controlled-voice board, and priced in the middle of the field after a 70% cut. If you need sound effects and ambience in the same pass, TTS-Next is a genuinely different capability rather than a tier of the same one.

Chinese-language and dialect coverage: Q​wen, without much argument. Twenty dialect regions and cross-language timbre transfer are not features G​PT-Live-1 has an answer to, and the preset asymmetry on O​penAI's side — Hindi and Korean get one voice each, both feminine — shows how differently the two vendors have prioritised language breadth.

English-first conversational product with an independent scoreboard behind it: G​PT-Live-1, with the caveat that you are buying a two-model system and the backend half is where the cost and the quality both live.

And if you are building on A​libaba's realtime tier specifically, the question to answer before you commit is the one no leaderboard will answer for you: what your backend costs per session. At an 85% cut the conversation itself became cheap. Whether the thing doing the thinking did is a separate invoice.

這些都不是假設：像這次發布這類版本背後那些有紀錄的 AI 失效與代理事件，全都收錄在OrcaRouter AI Incident Archive，其中每筆條目都附有自己的來源、嚴重程度、信心等級與爭議標記。