How much does Qwen-Audio-3.1-Realtime-Plus cost in Singapore?

In the Singapore region, audio input costs $6.40 per 1M tokens, while synthesized audio output costs $24.00 per 1M tokens. Text input is $0.80 and text output is $6.40 per 1M tokens.

What is the maximum context window for Qwen-Audio-3.1-Realtime-Plus?

The model supports a total context window of 262,144 tokens, with up to 245,760 tokens for input and a cap of 16,384 tokens for output generation.

Can I use Web Search and Function Calling simultaneously?

No, the model does not support enabling Web Search and Function Calling at the same time. You must execute these tasks sequentially in your application logic.

What are the audio format requirements for the API?

Input audio must be single-channel mono PCM at 16 kHz and 16-bit depth. Output audio is provided as single-channel mono PCM at 24 kHz and 16-bit depth.

What is the primary benefit of upgrading to version 3.1?

The primary improvement is background noise handling, which reduces false responses to background speech from 73.0% down to 13.0%, making it ideal for noisy environments.

Qwen-Audio-3.1-Realtime-Plus Pricing and Performance Guide

Qwen-Audio-3.1-Realtime-Plus Pricing and Specs. A macro close-up of a high-density liquid-cooled server blade rack, where vibrant blue coolant tubing weaves through tightly packed GPU modules, illuminated by a sharp, rhythmic strobe of — GPUniq

Data as of October 1, 20264 sources

Drafted with AI tools and edited by a human. The figures and conclusions were checked by a GPUniq editor.

TL;DR

Qwen-Audio-3.1-Realtime-Plus costs $6.40 per 1M audio input tokens and $24.00 per 1M audio output tokens in Singapore on Alibaba Cloud Model Studio. Context capacity reaches 262,144 tokens while false responses to background speech drop from 73.0% down to 13.0%. A full-duplex session constraint prevents Web Search and Function Calling from running simultaneously.

How Much Does the API Cost?

Alibaba Cloud Model Studio bills qwen-audio-3.1-realtime-plus by token modality and region. Audio processing uses separate rates for input ingestion and output synthesis.

Region	Text In ($/1M)	Audio In ($/1M)	Text Out ($/1M)	Audio Out ($/1M)
Singapore (SGP)	0.80	6.40	6.40	24.00
China (Beijing)	0.688	5.501	5.501	20.628

In Singapore, text input costs $0.80 per 1M tokens. Text output and audio input both cost $6.40 per 1M tokens. Synthesized audio output costs $24.00 per 1M tokens.

Deploying in Beijing drops rates to local pricing. Text input costs $0.688 per 1M tokens. Text output and audio input cost $5.501 per 1M tokens. Audio output costs $20.628 per 1M tokens.

Audio output drives total billing quickly. Live streams consume far more tokens than plain text responses.

How Does 3.1 Compare to 3.0?

Benchmark data shows performance gains over Qwen-Audio-3.0-Realtime across speech tasks.

Metric / Benchmark	Qwen-Audio-3.0-Realtime	Qwen-Audio-3.1-Realtime-Plus	Delta
Audio MultiChallenge	Baseline	Baseline + 5.09%	+5.09%
14-language QA (Big BENCH Audio)	Baseline	Baseline + 6.40%	+6.40%
τ²-Bench Audio Task Success	Baseline	Baseline + 3.59%	+3.59%
τ-Voice Adaptation (Speech-to-Text)	78.4%	82.0%	+3.60%
Background Speech False Response	73.0%	13.0%	-60.00%

Accuracy on the half-duplex speech-to-text adaptation of τ-Voice rose from 78.4% to 82.0%. Audio MultiChallenge improved by +5.09%, while the 14-language QA evaluation on Big BENCH Audio gained +6.40%.

The major operational fix is background noise handling. On Full-Duplex-Bench v1.5, false responses to background chatter dropped from 73.0% down to 13.0%.

What Is the Maximum Context Window?

qwen-audio-3.1-realtime-plus supports 262,144 tokens of total context.

The context splits unevenly between input and generation:

Maximum Input Allocation: Up to 245,760 tokens for audio streams, text prompts, and instructions.
Maximum Output Generation: Capped at 16,384 tokens per response.

The API automatically truncates history after approximately 50 turns or 300 seconds of continuous audio. Track frame duration client-side instead of assuming endless buffer retention.

Does It Support Search and Tools?

The model supports Function Calling and real-time Web Search. However, Web Search and Function Calling cannot both be enabled at the same time.

{
  "model": "qwen-audio-3.1-realtime-plus",
  "modalities": ["audio", "text"],
  "enable_web_search": true,
  "tools": [] 
}

If your stack requires external lookups alongside search, route them sequentially. Execute database lookups first using function calls, then pass that context into a search-enabled session.

The API offers system voices like longanqian_v3.1 and longanhuan_v3.1 alongside voice cloning. Note that setting an output voice language operates on a best-effort basis. The engine does not guarantee language overrides, and text output language remains unaffected.

What Are the Audio Stream Requirements?

You cannot download model weights. qwen-audio-3.1-realtime-plus is API-only, hosted on Alibaba Cloud Model Studio via WebSocket, AOQ, or WebRTC protocols.

Audio chunks must match exact specs to prevent errors:

Audio Input Format: Single-channel mono PCM, 16 kHz sample rate, 16-bit depth.
Audio Output Format: Single-channel mono PCM, 24 kHz sample rate, 16-bit depth.

Direction	Format	Sample Rate	Bit Depth	Channels
Client Input	PCM	16 kHz	16-bit	Mono
Stream Output	PCM	24 kHz	16-bit	Mono

Turn control uses push-to-talk, client VAD, or the native smart_turn auto-detection. Audio input quality may degrade over weak networks if format and sample rate requirements are not maintained.

Should You Upgrade to Version 3.1?

Upgrade if you deploy voice agents in noisy spaces. Reducing background false triggers from 73.0% to 13.0% fixes random unprompted interruptions.

Upgrade immediately for:

Noisy Environments: Warehouses, kiosks, or mobile apps with ambient noise.
Multilingual Coverage: Voice apps using English, Chinese, Japanese, Korean, German, French, Spanish, Portuguese, Russian, Indonesian, Italian, or Chinese dialects.

Hold off if:

Your backend executes function calls inside active web search sessions. You must refactor that workflow into separate calls before upgrading.

Methodology

Facts are compiled from the vendor's own announcement and documentation, linked below, at the time of writing; where the vendor hasn't published a number it is marked as such. GPUniq prices, when the model is already available here, come from our live catalog.