What is Qwen3.8-Omni-Flash?

Qwen3.8-Omni-Flash is Alibaba's omnimodal model released September 18, 2026. It accepts text, images, audio and video and returns text, with a 1M-token context window (991K max input, 131K max output on QwenCloud). It is built on the Qwen3.8-Flash-Next architecture and is the first Qwen omni model designed around agentic use: tool calling, web search, structured outputs, context caching and batch are all supported through OpenAI-compatible and DashScope APIs. Pricing is $0.15 per million input tokens and $0.47 per million output tokens, with cache hits at $0.016. It is API-only; no open weights were announced.

Is Qwen3.8-Omni-Flash cheaper than Gemini 3.8 Flash?

Yes, by 5x on input and 8x on output at list price. Qwen3.8-Omni-Flash is $0.15 input / $0.47 output per million tokens. Gemini 3.8 Flash is $0.75 / $3.75 as an introductory rate through December 31, 2026, rising to $1.50 / $7.50 on January 1, 2027, which would widen the gap to 10x and 16x. Both accept audio and video input. Real cost depends on tokens per task: Gemini 3.8 Flash is known to be token-hungry on Artificial Analysis's harness, and Qwen claims its agentic perception cuts long-video token use by 45.7%.

How does Qwen3.8-Omni-Flash compare on quality?

All figures are Alibaba's own as of September 20, 2026; independent results were not yet available. Qwen reports a 25%+ average gain over Qwen3.5-Omni-Plus across 29 evaluations, audio-visual reasoning close to Gemini 3.8 Flash, and overall audio performance above Gemini 3.8 Flash. On OmniVideoBench its agentic perception raised accuracy from 63.4 to 67.8 while cutting tokens from 145,736 to 79,117. Gemini 3.8 Flash scores 58.7 on the Artificial Analysis Intelligence Index and remains the more proven general-purpose model.

What are the input limits for Qwen3.8-Omni-Flash?

Per Alibaba Cloud Model Studio documentation: video files up to 2 hours and 2 GB by URL, audio files up to 3 hours, audio input in 113 languages and dialects, video sampled at up to 15 fps, and two-channel stereo or four-channel FOA spatial audio via use_multichannel. Maximum reasoning length is 262K tokens, with thinking on by default at xhigh reasoning effort. It is served from six regions: Beijing, Singapore, Hong Kong, Tokyo, Frankfurt and Virginia.

Quick Answer

Qwen3.8-Omni-Flash vs Gemini 3.8 Flash: Omnimodal API (2026)

Published: September 20, 2026

Side by side

	Qwen3.8-Omni-Flash	Gemini 3.8 Flash
Maker	Alibaba Qwen	Google DeepMind
Released	September 18, 2026	September 2, 2026
Inputs	Text, image, audio, video	Text, image, audio, video
Output	Text only (use Qwen3.5-Omni for speech)	Text
Context	1M (991K in / 131K out)	1M (64K max output)
Price per MTok	$0.15 in / $0.47 out	$0.75 / $3.75 intro → $1.50 / $7.50 from Jan 1, 2027
Cached input	$0.016	Lower tier available
Max video	2 hours / 2 GB by URL	Long video supported
Max audio	3 hours, 113 languages	Multilingual
Open weights	No (base Qwen3.8-Flash-Next is open)	No
Regions	6 incl. Frankfurt, Virginia	Global
Knowledge cutoff	Not stated	March 2026

Verified September 20, 2026 against the Qwen blog, Alibaba Cloud Model Studio docs and Google’s Gemini pricing page.

What Qwen3.8-Omni-Flash is

Alibaba’s Qwen team calls Qwen3.8-Omni-Flash “its first omni-modal model built around agentic capabilities.” The pitch is a single loop — understand the media, plan the task, execute with tools, deliver a result — rather than a captioning model bolted onto an LLM. Concretely:

Native audio-video understanding in one model, with text output. Speech generation is deliberately left to Qwen3.5-Omni.
Agentic perception for long video. Instead of reading a two-hour file start to finish, the model starts from the question, decides which segments to watch and listen to, and gathers evidence in coarse-to-fine rounds. Qwen reports OmniVideoBench accuracy rising from 63.4 to 67.8 while tokens fell 45.7%.
Full agent plumbing: function calling, web search, structured outputs, context caching and batch calls, via both DashScope and OpenAI-compatible Chat Completions/Responses APIs.
Open tooling around a closed model. Qwen open-sourced Qwen-MM-Plugins (Apache-2.0), which installs omni capabilities as Skills plus optional MCP servers into Claude Code, Codex, Gemini CLI, OpenClaw and others — for example omni-video2note turning a tutorial into an illustrated PDF.

Reported gains over Qwen3.5-Omni-Plus: 25%+ average across 29 evaluations, +36.5 points on WildClawBench-MM, +22.3 on AgenticVBench. All vendor numbers; no independent evaluation had been published by September 20.

What Gemini 3.8 Flash is

Google’s Gemini 3.8 Flash (September 2, 2026) is the default model in AI Mode and the workhorse of the Gemini API: 1M context, 64K max output, text/image/audio/video in, knowledge cutoff March 2026. It scores 58.7 on the Artificial Analysis Intelligence Index and, on the same harness, is the most token-hungry frontier model — about 123M output tokens to complete the full index versus ~16M for GPT-6 Astra. Its price is an introductory $0.75 / $3.75 per million tokens through December 31, 2026, doubling to $1.50 / $7.50 on January 1, 2027. A separate gated Gemini 3.8 Flash Cyber variant exists under the Fairwind Program and is not on the standard API. See Gemini 3.8 Flash vs 3.7 Flash vs Opus 5.

Where they differ in practice

Cost per hour of media. Qwen says audio input costs over 98% less per hour than Qwen3.5-Omni-Plus and audio-visual input over 93% less. Against Gemini the list-price gap is 5x input / 8x output today and 10x / 16x after January 1, 2027. Because Gemini 3.8 Flash burns more tokens per task, the real gap on agentic video work is likely larger than list price suggests — but only your own transcript will prove it.

Long video. Qwen’s agentic perception is the differentiator: it is the only one of the two that publishes a token-reduction mechanism for long video (145,736 → 79,117 tokens on OmniVideoBench). Gemini processes video linearly.

Agent harness fit. Qwen-MM-Plugins makes Qwen3.8-Omni-Flash the easiest omni model to bolt onto a coding agent today. Gemini’s equivalent is Gemini CLI, which Google deprecated in September 2026 after the Plugin4Shell disclosure (see what is Plugin4Shell).

Data residency. Qwen serves from Frankfurt and Virginia as well as four Asian regions, which answers the usual objection to Chinese-hosted APIs for EU and US teams; it does not answer procurement policies that exclude Alibaba Cloud outright.

Trust and track record. Gemini 3.8 Flash has independent scores and Google’s SLAs. Qwen3.8-Omni-Flash has vendor benchmarks and a two-day-old API.

Recommendation

Bulk audio/video understanding at scale (call analysis, lecture indexing, video QA): trial Qwen3.8-Omni-Flash first; the price gap is too large to ignore and the region list covers EU/US.
General assistant or search-grounded product where the media is a minority of traffic: stay on Gemini 3.8 Flash for now, and re-price before January 1, 2027 when its intro rate ends.
Spoken output required: neither — Qwen3.5-Omni or a dedicated TTS API (see best TTS APIs 2026).

Side by side

What Qwen3.8-Omni-Flash is

What Gemini 3.8 Flash is

Where they differ in practice

Recommendation

Sources