Qwen3.8-Omni-Flash cuts audio input costs by 98%

#llm #api #moe #china

Alibaba released Qwen3.8-Omni-Flash on September 18, 2026, a model that accepts text, images, audio, and video and replies in text. The price is the headline. Compared with Qwen3.5-Omni-Plus, audio input costs more than 98% less per hour, audio-visual input more than 93% less, and video input about 89% less, according to MarkTechPost and AlphaSignal.

Omnimodal means one model handles several input types natively, rather than a pipeline that transcribes audio first and then reads the transcript. Text costs $0.15 per million input tokens and $0.47 per million output tokens, with cached input at $0.016 per million.

What it takes in

Limit	Value
Context window	1M tokens, up to 991K input and 131K output
Reasoning length	Up to 262K tokens
Video	Up to 2 hours, 2 GB, sampled up to 15fps
Audio	Up to 3 hours
Languages	113

Availability is the catch for anyone who liked the last release. This one is API-only, through QwenCloud, Alibaba Cloud Model Studio, and Qwen Studio, with no open weights announced. That is a change of posture from Qwen3.8-Flash-Next, which shipped open weights in August 2026 and supplies the architecture here.

That architecture is a mixture of experts. The network holds about 125 billion parameters across 512 experts but activates only around 6 billion for each token, alongside a 51 billion-parameter embedding table. The design is why a model this large can be priced like a small one.

Agentic perception, and the token savings

The more interesting claim is about how the model watches video. Rather than sampling frames at a fixed interval, it inspects the source in two passes: a coarse scan to find the relevant stretch, then a closer look at that stretch. AlphaSignal describes it as agentic perception, and quotes Qwen calling this "the first model in its family to make omnimodal perception part of an agent workflow."

The reported savings differ by benchmark, so both numbers are worth stating. MarkTechPost reports one task dropping from 145,736 to 79,117 tokens, a 45.7% reduction. AlphaSignal cites 51.8% fewer tokens on OmniVideoBench, with an average agent gain of 19.5 points over the previous model. BenchLM, which tracks published benchmark rows, notes it has results for only 19 of its 446 benchmarks so far and publishes no overall score yet.

Alibaba also released Qwen-MM-Plugins under the Apache 2.0 license. The toolkit adds memory, video-to-note, and editing capabilities, and is meant to plug multimodal input into existing agent frameworks such as Claude Code and Gemini CLI. Extended thinking is on by default at the highest setting, and can be turned off.

What this means for developers

Price changes of this size change what is worth building. Transcribing a support call, summarizing a recorded meeting, or scanning security footage used to be projects where the model bill decided the design. At under a dollar for a large batch of text tokens, and with audio down 98%, the calculation moves toward simply sending the media.

Check the tokenizer math before you trust a quote, though. Audio and video consume tokens at rates that depend on length, frame rate, and how much of the source the model decides to inspect closely. Agentic perception means the token count varies with content, so run your own worst case rather than an average.

The API-only release is the real constraint for teams with rules about where data goes. Audio and video are exactly the inputs that carry personal information, and this model runs on Alibaba's cloud. If your last Qwen deployment relied on local weights, this is not a drop-in successor. Benchmark coverage is still thin, so treat the agent numbers as vendor-reported until independent evaluations fill in.

This article was first published on Tech AI Wire.

Also available in

Deutsch · 日本語 · Français · Español · Português

Related on Tech AI Wire

Qwen3.8-Flash-Next holds 125B parameters but runs only 6B

DEV Community