Qwen3.8-Omni-Flash vs Gemini 3.8 Flash

This is the comparison Alibaba picked itself. Qwen3.8-Omni-Flash, released September 18, 2026, claims audio-visual performance close to Gemini 3.8 Flash and overall audio performance above it, so the question is whether a five times cheaper rate card buys you the same job. Qwen bills $0.15 per million input tokens and $0.47 output with cache hits at $0.016. Gemini 3.8 Flash, released September 2, bills $0.75 and $3.75, and Google printed the expiry in a launch-post footnote: both sides double to $1.50 and $7.50 on January 1, 2027. So the gap today is five times on input and eight times on output, and it widens to ten and sixteen after the new year. Both carry roughly 1 million token context windows and both take text, image, audio, and video. The differences are in what comes out and who says so. Qwen returns text only, so speech generation needs a separate model, while Gemini handles more output paths on one endpoint. Qwen has published no independent benchmark at all: the 29-evaluation average, the 36.5 point WildClawBench-MM gain, and the OmniVideoBench lift from 63.4 to 67.8 are all vendor-reported with Qwen's own estimation method, whereas Artificial Analysis publishes an index and a cost-per-task figure for Gemini 3.8 Flash. Qwen's genuine architectural claim is agentic perception: it decides which parts of a long video to watch rather than reading it front to back, which it says cuts token use about 45.7 percent while raising accuracy. If that holds on your footage, the effective gap is wider than the per-token gap. There is one more axis that is not about quality. Gemini runs on Vertex and AI Studio with the compliance surface that implies; Qwen3.8-Omni-Flash runs on Alibaba Cloud Model Studio in six regions, and Alibaba was one of six labs named in the September 8, 2026 joint CISA, NSA, and FBI advisory on model distillation. That does not change a benchmark, but it will change some procurement reviews.