Qwen-Audio 3.1 cuts voice API prices up to 95%

The cheapest part of a voice agent used to be the part you did not build. That is changing fast. Alibaba's Qwen team shipped Qwen-Audio 3.1 — five models covering recognition, synthesis and realtime conversation — and cut prices across the lineup by up to 95%. If you are paying a US vendor per minute of speech, the Qwen-Audio 3.1 price cut is the number your next renewal negotiation should start from.

What actually happened

Per The Decoder, the September 23 release upgrades the three existing models — ASR, TTS and Realtime — and adds two new ones:

• ASR-Next for audio understanding: multi-speaker labeling with timestamps, emotion detection, and recognition of ambient and machine sounds.
• TTS-Next for audio creation: voice, sound effects and background audio generated in a single pass, aimed at audiobooks, podcasts, games and ads.

The price moves, consistent across independent reports: ASR down up to 95%, TTS down roughly 70%, realtime voice down roughly 85%. 36Kr published per-model list pricing in yuan — ASR-Flash at 0.8 in / 2.7 out per million tokens, TTS-Flash at 1.5 / 12, TTS-Next at 6 / 12, and Realtime-Plus ranging 5–40 in / 40–150 out by tier. Those are Alibaba Cloud list prices in China; convert and confirm against the region you would actually deploy in before you put them in a spreadsheet.

Capability moved too, not just price. The recognition model handles 30 languages and 16 Chinese dialects and strips filler words and repetitions automatically — the post-processing step most transcription pipelines currently pay a second vendor for.

Why cheap speech models matter for your business

Price cuts of this size reprice the whole category, not just one vendor. You do not need to deploy Qwen for this release to matter. You need to know the number exists when your incumbent's renewal lands. A 70% cut on synthesis from a credible competitor is leverage, and it costs you one email to use.

Falling ASR cost changes what is worth transcribing. At old prices, you transcribed the calls someone flagged. At a twentieth of that, you transcribe everything — every support call, every voicemail, every sales conversation — and the transcript becomes a searchable asset rather than a line item. Speaker labels with timestamps are what make that corpus queryable instead of a wall of text.

Do not swap vendors on price alone. Alibaba Cloud is the deployment path, and that carries real questions: where audio is processed, what your customer contracts say about data location, and what your industry's rules require. For a Phoenix services business recording customer calls, that is a conversation with your counsel before it is a conversation with your CTO. Answer it first; the discount is still there afterward.

Build the seam now. Put speech behind one internal interface — transcribe(audio) -> text, speak(text) -> audio — with the provider as configuration. We do this on every voice build, and the reason is exactly this week: the pricing under you moves by an order of magnitude and you want that to be a config change, not a refactor.

Sources: The Decoder, 36Kr.