Qwen3.8-Omni-Flash Hits Alibaba's API With 1M Context and a Claimed 98% Audio Price Cut

Our verdict: Conditional. Qwen3.8-Omni-Flash is a real, callable omnimodal model with a 1M-token context, and Qwen says audio input is now more than 98% cheaper per hour. That claim could reprice transcription and meeting-summary work. But every benchmark is Qwen's own, the model answers in text only, and we could not read a rate in either source we cite: Qwen's launch post shows its rates only as an image, and Model Studio's documentation lists no price. Test it, and read the rate in your own console before you move a workload.

What happened

Alibaba's Qwen team released Qwen3.8-Omni-Flash (Qwen's blog post is dated 2026-09-14). It is callable now on Alibaba Cloud Model Studio as qwen3.8-omni-flash.

• Input: text, images, audio and video.
• Output: text only.
• Context: 1M tokens.
• Mode: Model Studio describes it as a non-realtime model for audio and video understanding, meeting summaries and content analysis. It supports function calling and web search, and thinking mode is set with the reasoning_effort parameter (default xhigh).

Qwen's launch claims:

• Audio input more than 98% cheaper per hour, and audio-visual input more than 93% cheaper.
• An average score more than 25% higher than Qwen3.5-Omni-Plus across 29 evaluations.
• Audio-visual performance "close to Gemini 3.8 Flash", and overall audio performance that exceeds it.
• 69.6 on UniClawBench, a multimodal tool-use benchmark.

All of these numbers are Qwen's own evaluations and Qwen's own pricing statements.

Why it matters

Audio is where long inputs get expensive. If the per-hour cut holds, hours of recorded meetings or calls become cheap enough to process in bulk with one model that also reads slides and video. That is why we rate this Conditional rather than ignoring it.

The catch is evidence. We have no independent benchmark, and neither source we cite gives a readable price. Qwen picked Gemini 3.8 Flash, which we rate Recommended, as its comparison, so that is the model to beat on your own samples.

What changes for you

• Where to get it: the Model Studio API as qwen3.8-omni-flash, listed in China (Beijing), Singapore, China (Hong Kong), Japan (Tokyo), Germany (Frankfurt) and US (Virginia). Qwen's launch post also names a separate Qwen3.8-Omni-Flash-Realtime model for realtime use and the Qianwen AI platform.
• Spoken replies: this model returns text. If your agent needs to talk back, plan for the Realtime variant Qwen announced, or a separate speech model.
• Self-hosting: we found no open weights for the model itself, and Model Studio offers it as a hosted API. Qwen's launch post says it open-sourced companion tools on GitHub, including Qwen-Live Harness (a realtime runtime) and Qwen-MM-Plugins.
• Before you switch: confirm the rate in your Model Studio console, then run your own audio and video samples against Gemini 3.8 Flash.