Alibaba's Qwen team released Qwen3.8-Omni-Flash on September 18 -- a native omni-modal model that reads text, images, audio and video in a single 1-million-token context window (991K input, 131K output, plus a 262K reasoning budget) and calls tools to act on what it finds. It's available only through Alibaba's Qianwen platform and Model Studio API, across six regions (Beijing, Singapore, Hong Kong, Tokyo, Frankfurt, Virginia); no open weights shipped alongside it, and international pricing is $0.15 per million input tokens and $0.47 per million output tokens, with cached input at $0.016.
On Alibaba's own numbers, the model improved by more than 26% on average across 30 internal evaluations compared with its predecessor, Qwen3.5-Omni-Plus, with the clearest gains on agentic video tasks (AgenticVBench: 36.8, up 22.3) and long-audio comprehension (LongAudioSpan: 82.7, up 8.3). Against Google's Gemini 3.8 Flash, Alibaba's own benchmarks show a genuine split rather than a clean win: Qwen leads decisively on audio-centric tasks (SpotSoundBench: 67.2 vs. 39.7) and trails on pure video reasoning (AgenticVBench: 36.8 vs. 45.0).
The launch's headline cost claim is softer than the percentage suggests. Alibaba says the new model costs 98% less per hour of audio input than its predecessor, and 93% less per hour of combined audio and video -- but by the company's own methodology, that's a per-hour figure extrapolated from a 2-minute sample multiplied by 30, and the audio-video comparison specifically assumes 1-frame-per-second video sampling, a far lower capture rate than most real video-analysis workloads would use. A real price cut is buried inside a comparison that also quietly changes what's being measured.
What Alibaba's "98% cheaper" claim includes
- $0.15 / $0.47 · per 1M tokens, in/out
- Official international API price
- 98% · cheaper audio (claimed)
- Alibaba's per-hour comparison vs. Qwen3.5-Omni-Plus
Includes: A real per-token price cut, extrapolated to a per-hour figure from a 2-minute sample multiplied by 30
Excludes: Real-world usage patterns at full sampling rates; the 93%-cheaper audio+video figure additionally assumes 1-frame-per-second video capture - 36.8 vs. 45.0 · AgenticVBench score
- Qwen3.8-Omni-Flash vs. Gemini 3.8 Flash
As with Alibaba's Qwen3.8-Max launch earlier this year, no independent benchmark exists yet for Qwen3.8-Omni-Flash -- every number above is the company's own, run on its own evaluation suite against its own choice of comparison points.
- Alibaba released Qwen3.8-Omni-Flash on Sept. 18, a 1-million-token omni-modal model.
- It handles text, images, audio and video, priced at $0.15/$0.47 per million tokens internationally.
- It leads Gemini 3.8 Flash on audio benchmarks, trails it on pure video reasoning.
- Alibaba claims 98% cheaper audio processing than predecessor Qwen3.5-Omni-Plus.
- Caveat: that 98% figure blends a real price cut with a lower video sampling rate; no independent score exists yet.