A generated title card reading "Which tier are you actually calling?" above the line "Qwen-Audio-3.1 is a family of nine model ids, not one endpoint", with nine chips in three rows of three reading tts-plus, tts-flash, tts-next, realtime-plus, realtime-flash, asr-flash, asr-flash-streaming, asr-flash-filetrans and asr-flash-message, a caption "Listed in stages, 20 September to 27 October 2026", and a footer citing Alibaba Cloud Model Studio. The OrcaRouter logo is composited bottom-right.
Guides & Insights

Qwen-Audio-3.1-TTS-Plus vs Qwen-Audio-3.1: One Is a Model, the Other Is a Badge

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

The short answer: you cannot call Q​wen-Audio-3.1, because there is no such model. There is a Q​wen-Audio-3.1 family — nine model IDs that A​libaba Cloud Model Studio documents separately, on different release dates, with different prices and different voice lists — and Q​wen-Audio-3.1-TTS-Plus is the premium member of it, listed in A​libaba's release notes on 21 October 2026, five weeks after the family's first ID appeared on 20 September. The sibling most people actually mean when they write "Q​wen-Audio-3.1" in a support ticket is Q​wen-Audio-3.1-TTS-Flash, the general-purpose text-to-speech tier. Those two are the same job at different prices. Everything below is about choosing between them, and about the one place where the expensive one cannot do what the cheap one can.

A screenshot of Alibaba Cloud Model Studio's model lifecycle and updates changelog for the Singapore region, showing the page's dated release-note rows for October 2026 alongside the changelog navigation.

First, the nine IDs

A​libaba's own release notes are the clearest source, and they contradict the impression you get from a news story. There was no single launch day. The family arrived in pieces over five weeks, and the pieces are still arriving.

• q​wen-audio-3.1-realtime-plus — listed 20 September 2026 in both Singapore and Beijing. Duplex speech-to-speech conversation, with voice cloning.

• q​wen-audio-3.1-asr-flash, together with its -streaming, -filetrans and -message variants — listed 21 September 2026. Four recognition endpoints, not one: short-form, streaming, offline file transcription, and an optimised voice-message variant.

• q​wen-audio-3.1-tts-next — listed 21 September 2026 in Beijing, 22 October 2026 in Singapore. A​libaba describes it as an "AudioGen" model: speech, sound effects and ambience generated in a single pass from text, timestamps and reference audio. It is not a TTS tier in the ordinary sense.

• q​wen-audio-3.1-tts-flash — listed 12 October 2026 in both regions. Streaming speech synthesis for assistants, real-time conversation and customer service.

• q​wen-audio-3.1-tts-plus — listed 21 October 2026 in both regions. High-quality synthesis for premium voices and content creation, with voice cloning described as more robust to noisy reference audio.

• q​wen-audio-3.1-realtime-flash — listed 27 October 2026. The cheaper duplex tier, with a 262,144-token context window, function calling and web search.

Read those dates against each other and the framing problem is obvious. A "Q​wen-Audio-3.1 launch" in late September covered the recognition and realtime tiers; the two text-to-speech tiers that most articles are actually about did not appear until October. If you built against the family in the week of the announcement, the Plus tier was not in it yet.

A generated two-column scoreboard headed "Plus vs Flash". The qwen-audio-3.1-tts-plus column reads voices four flagship, languages narrow catalogue, billing per character input only, rate $0.30 per 10k characters in Singapore, output audio not charged, independent score 1,179 Elo rank 2 of 49. The qwen-audio-3.1-tts-flash column reads voices full catalogue with 15 accented English, languages 8 Chinese dialects plus 15 more, billing per token input and output, rate $0.23 in and $1.87 out per 1M tokens, output audio billed at 8x input, independent score none published. The OrcaRouter logo is composited bottom-right.

The comparison nobody runs, and the one that decides it

Plus against Flash is the only like-for-like pair inside the family. Both take text, return audio, bill through the same account, and sit behind the same published ceiling. The differences are not the ones the tier words imply.

• Voices — q​wen-audio-3.1-tts-plus has four. A​libaba publishes exactly four flagship system voices for it: longanlingxin_v3.1, xunanchuan_v3.1, longanhuan_v3.1 and longanfengyue_v3.1, three female and one male. q​wen-audio-3.1-tts-flash has the catalogue: multilingual voices covering eight Chinese dialects plus Japanese, Korean, French, German, Portuguese, Italian, Vietnamese and Indonesian; a long list of Mandarin premium voices; additional-language voices for Icelandic, Maltese, Swahili, Kazakh, Ukrainian and more; and fifteen English voices with named accents — Emily, Luna, Eric and Luca as British, and Abby, Annie, Ava, Beth, Betty, Cally, Cindy, Donna, Andy, Brian and David as American.

• Billing unit — Plus is billed per input character with no charge for output. Flash is billed per token, input and output separately, at $0.23 and $1.87 per million tokens in Singapore and $0.212 and $1.696 in Beijing.

• Rate card — Plus is $0.30 per 10,000 characters in Singapore and $0.20 in Beijing. Flash has no per-character figure, because its unit is the token.

• Character counting, if you are on Plus — a Chinese, Japanese or Korean character counts as two; a Latin letter, digit, punctuation mark or space counts as one. SSML tags are not counted. 你好 is four characters.

• Rate limit — 180 requests per minute for both, in both regions. The ceiling does not separate them.

• Output — PCM, WAV, MP3 or O​pus, sample rates up to 48 kHz, on both. Non-streaming requests return a URL valid for 24 hours; streaming returns chunks. A connection closes if no new task starts within 60 seconds of the previous one ending.

• Free quota — Plus and Flash each carry 1,000,000 free characters or tokens respectively in Singapore, valid for 90 days from activation, model release or application approval.

The line that matters is the first one, and it cuts against the naming. The tier called Plus has four voices. The tier called Flash has the voices, including the only British-accented English voices A​libaba publishes anywhere in this family.

The measured model is not the model that can speak British English

Q​wen-Audio-3.1-TTS-Plus is the only member of the family with an independent score, and it has three. On Artificial Analysis's Controlled Voice arena, where every entrant synthesises the same eight cloned reference voices, it ranks 1,179.08 Elo, second of 49, behind HeyGen Voice at 1,188.52 and ahead of HeyGen Voice Turbo at 1,172.56 — the top two intervals overlap, so second and first are not statistically separated on that board. Split by accent on 7 November, it is first of 49 in US English at 1,207.07 and fifth in UK English at 1,136.72. On the Provider Voice arena, where each vendor supplies its own native voices, it is third of 104 at 1,303, behind Eleven v4 at 1,339 and Eleven v4 Turbo at 1,335.

Now put that beside the voice list. A team that reads "first in US English" and orders the Plus tier is buying a four-voice model. If the product needs a British female voice — the single most common requirement in English-language consumer voice work — the Plus tier does not have one. Flash does, four of them. And Flash has no score on either arena, so there is nothing to compare with the 1,207.

That is the trap in this matchup, and it is structural rather than a defect. A​libaba entered one tier in the arena. The tier it entered is the premium one, because that is the one worth benchmarking. The consequences for a buyer are not symmetric: the model with the number cannot do the accent, and the model that can do the accent has no number. Choosing on rank alone gets this backwards.

Costing it, honestly

Every comparison of these two tries to subtract one price from the other. That subtraction is not available. Plus is denominated in characters; Flash in tokens, with the audio counted on the output side. There is no published character-to-token ratio for this model, so any single "Flash is X% cheaper" claim is invented.

What can be said is where the risk sits. Flash's output rate is eight times its input rate in Singapore and eight times in Beijing, so the audio dominates a Flash invoice on any clip of realistic length, and the number of output tokens depends on how long the model decides to speak. Plus has no output charge at all: the invoice is a function of the script you sent, and nothing else. For fixed scripts of known length, model both and measure — the free quota covers it. For anything where pacing is the model's choice, Plus's flat per-character price is worth paying for predictability alone, and that argument is independent of quality.

One structural warning about the public numbers. The Controlled Voice arena lists the Plus tier at $19.31 per million characters. A​libaba's billing page lists $30.00 per million in Singapore and $20.00 in Beijing. The arena figure is lower than both and closely matches the previous generation's Beijing rate, q​wen-audio-3.0-tts-plus at $0.19253 per 10,000 characters. We cannot tell from outside whether that is a stale mapping or an unpublished rate. Budget from the region you actually call.

A screenshot of the OrcaRouter models catalogue showing the "226 models, 17 providers, one API key, one bill" header, the filter rail for input modalities, context length, price, status and access tier, and the first model cards with a copyable model ID.

How the choice resolves in a pipeline

In practice most teams need both tiers and the decision becomes operational rather than commercial. A worked shape: the voice request carries a locale and a voice ID; the voice ID determines the tier, because the catalogue does; the tier determines the endpoint; and a failover policy decides what happens when the chosen endpoint degrades halfway through a batch of forty thousand clips.

That is a routing problem, and it is the layer OrcaRouter is built for — a routing DSL that expresses the voice-to-tier mapping as configuration rather than as branching code, automatic failover so a failing upstream hands the request to a second engine instead of losing the clip, and one API key across 200-plus models so the Q​wen speech leg and whatever you fuse with it sit behind a single credential and a single price basis, with provider list price passed through at 0% markup. The boundary is worth stating plainly: OrcaRouter does not host A​libaba's speech models. The Q​wen legs of that rule are the vendor's own API. What we supply is the key, the rule and the failover.

The checklist before you spend anything

• Confirm the ID, not the tier word. Plus means fewer voices and character billing. Flash means the catalogue and token billing. Next means sound scenes. None of those words describes what the model does.

• Check the voice list before the rate card. A voice you do not have is not a price, and the accent you need may only exist on the tier you were about to skip.

• Check the region. Singapore and Beijing differ on every price in this family, and the English premium voices are the one category A​libaba documents as available in both.

• Treat a leaderboard rank as a rank for one member. Q​wen-Audio-3.1-TTS-Plus at 1,179.08 says nothing about q​wen-audio-3.1-tts-flash, which is on neither arena, and nothing about q​wen-audio-3.1-tts-next, which is a different product category with the same prefix.

• Ask what "announced 23 September" meant for you. For the TTS tiers it meant nothing: they were listed in October.

Sources, separated by who supplied them

• Artificial Analysis, read 9 November 2026 — Controlled Voice: Q​wen-Audio-3.1-TTS-Plus 1,179.08 Elo, #2 of 49, interval 1,166.08–1,192.08; US English 1,207.07, #1 of 49, 1,190.07–1,224.07; UK English 1,136.72, #5 of 49, 1,118.72–1,154.72; HeyGen Voice 1,188.52 pooled, 1,187.28 US, 1,181.55 UK. Provider Voice: Q​wen-Audio-3.1-TTS-Plus 1,303, #3 of 104. The 7 November accent-split publication date is Artificial Analysis's.

• A​libaba Cloud Model Studio, read 9 November 2026 — per-model release dates from the model-lifecycle and updates page; prices, billing units, character-counting rules and free quotas from the billing page; 180 RPM from the rate-limits page; the four Plus flagship voices and the Flash voice catalogue, including the fifteen accented English voices and 51 additional-language voices, from the Q​wen-Audio-TTS voice list page.

• Absent from every source above — any open-weights release for the 3.1 speech family. Q​wen's public audio repositories stop at Qwen2-Audio, July 2024, and the arena records Q​wen-Audio-3.1-TTS-Plus as closed-weight. There is no checkpoint to download and no self-host path.

• Ours — the observation that the arena's price column sits below both published vendor rates, the argument that a character-to-token conversion cannot be made, and the claim that the measured tier is the one without British English voices. The first is unresolved, the second rests on the absence of a published ratio, and the third follows from the vendor's own two voice lists.