Quick answer
Alibaba released Qwen3.8-Omni-Flash on September 18, 2026: a natively omnimodal model that accepts text, images, audio, and video as input, with a one-million-token context window, priced at $0.15 per million input tokens and $0.47 per million output tokens (cache hits $0.016). Alibaba reports a gain of more than 25% over the previous Qwen3.5-Omni-Plus on its aggregate benchmarks. Separately, Qwen 3.8 Max became the first Alibaba model to reach the top three on the WebDev Arena leaderboard this month.
"Omnimodal" means one model handles every input type natively, rather than stitching a vision model, a speech model, and a text model together behind an API. The practical difference is that the model can reason across modalities — describe what is happening in a video while transcribing its audio and answering a typed question about both — without hand-offs that lose information. Until this month that capability cost frontier prices. Qwen3.8-Omni-Flash offers it for less than a fifth of the cost of Gemini 3.8 Flash.
What it does
- Accepts text, image, audio, and video input in a single request
- One-million-token context — enough for hours of transcribed audio or long videos
- Text output, with speech output available through Alibaba's companion services
- Available through Alibaba Cloud Model Studio and third-party gateways
- Part of the Qwen3.8 family, whose Flash variant shipped in late August
Why the price matters more than the benchmark
At $0.15 per million input tokens, processing an hour of video transcript costs a fraction of a cent. That changes which products are viable: always-on meeting analysis, video search across an archive, accessibility tools that describe the world in real time, and customer-service systems that watch a screen-share. None of these were economical at 2025 multimodal prices. The 25% improvement over the previous Omni model is welcome, but the pricing is what will get it deployed.
The caveats
- Data residency and compliance: Alibaba Cloud hosting is a non-starter for some Western enterprises, though open-weight Qwen models can be self-hosted
- Benchmark claims are Alibaba's own; independent multimodal evaluations lag text ones by months
- Chinese-model content policies differ from Western vendors' on politically sensitive topics
- Omnimodal quality on Western-accented speech and non-Chinese video content is less documented
Two years ago the question was whether Chinese labs could match the frontier. This month Qwen 3.8 Max reached the top three on a Western leaderboard while Qwen3.8-Omni-Flash undercut everyone on price. The question now is whether Western buyers will use them.
Related reading
Bottom line
Qwen3.8-Omni-Flash is the cheapest way to give software eyes and ears in 2026. For developers outside regulated industries, it is worth a serious pilot. For everyone else, it is the price point Google, OpenAI, and Anthropic will have to answer.
