Every Qwen3.8-Omni-Flash workflow ends in text your own tools have to execute

Call it through Chat Completions or the Responses API and what comes back is text ^[3]. Ask it to cut down a 40-minute interview and it can analyze scenes, speakers and events, then invoke a custom tool through function calling ^[5]. The encode, the render and the delivery are your code. Alibaba ships the open-source Qwen-MM-Plugins repository as a tool layer for multimodal agents, and those tools sit outside the model's documented native output ^[6]. If you already own an encoder and a renderer, the planner slot above them is the cheap thing to try.

The published international rates are $0.15 per million input tokens, $0.016 per million cache-hit input tokens and $0.47 per million output tokens ^[8]. A cache hit is therefore about 9.4 times cheaper than a fresh input token ^[1], and an output token costs about 3.1 times an input token ^[2]. On a single chatbot turn that spread is noise. On an agent that walks the same recording eight times, passes two through eight cost roughly a tenth of the first, as long as the session stays alive and the cache actually hits. Alibaba applies modality-specific token-conversion rules while publishing one input rate for the model instead of separate rates for text and media ^[9].

Alibaba says its estimated API cost per hour of audio input fell by more than 98%, and audiovisual input by more than 93% ^[11]. The method behind those percentages: price two minutes of source material, multiply by 30, and use 720p video sampled at one frame per second for the audiovisual case ^[12]. Two minutes times 30 is where the hour comes from ^[3], and one frame per second across an hour is 3,600 sampled frames ^[4]. For that number to transfer, your footage has to be 720p and one frame per second has to be enough for the decision you are asking the model to make. Independent testing has not validated the comparison ^[13].

The quality figures are also Alibaba's own. It reports an average improvement of more than 25% over Qwen3.5-Omni-Plus across 29 evaluations, gains of 36.5 points on WildClawBench-MM and 22.3 points on AgenticVBench, and a 69.6 score on UniClawBench ^[14][15]. Alibaba says audio performance exceeded Gemini 3.8 Flash overall and audiovisual performance came close to Google's model ^[16]. Those are selected, company-reported results from the launch announcement, not independent benchmarks ^[17].

What the model accepts does not change what you have to build around it. Long media files still leave developers with engineering work around storage, processing and token budgets ^[18]. Alibaba has not published the parameter count, training-compute budget or training-data composition in its current model documentation ^[19]. The launch announcement is still marked draft ^[20]. The production documentation was updated on September 18th and lists Qwen3.8-Omni-Flash as available in Beijing, Singapore, Hong Kong, Tokyo, Frankfurt and Virginia ^[21].