qwen3.8-omni-flash vs gemini3.8-flash we asked to explain the memes: rickroll, distracted boyfriend, disaster girl. gemini won 2 to 1

the contenders: @Alibaba_Qwen's new native omni model, audio + video in, tested in qwen studio, thinking mode – against @GeminiApp's current flash via @OpenRouter, $0.75/m in, $3.75/m out

the job: 6 memes as an image or an 8-second clip – what happens, quote the speech or name the song, why it's funny, origin, confidence 0–10. test 7: a 6m 25s video with no hint – who, what show, what year, where; who said what in 02:07–03:15; first "healthcare dot gov" and how often; every topic with a timestamp. both worked out it was between two ferns with obama. every score is ours, by hand.

- rounds won, 7 total (4 ties)
#1 gemini – 2
#2 qwen – 1

- facts right, memes 1–6, out of 30
#1 gemini – 20, 1 made up
#2 qwen – 13, 3 made up

- answer time per meme
#1 gemini – 7s
#2 qwen – 68s, up to 3m 22s

- test 7, a tie
gemini – 17/17 speakers, year 2014 ✓, 19 topics, 27s
qwen – 17/17 speakers, 21 topics, year 2013 ✗, 4m 50s

- cost per request, file + prompt + answer
#1 qwen – ~$0.001, estimated at list price, the chat shows no tokens
#2 gemini – $0.0025, billed by openrouter

observations:
• test 4, numa numa: qwen studio strips the soundtrack – "i don't have audio, only frames", says qwen's thinking. 8 frames plus an asr transcript of any speech go in, music never does – so the song came back as "silent frames". fed the same 8 seconds as an mp3, qwen named dragostea din tei by o-zone in 3.5s.
• speech reaches qwen as text: "you out charlie" fixed to "ouch, charlie" from memory. on duang, qwen got jackie chan's "very oily" hair right, gemini heard "soft".
• qwen rates itself 8–10 on every meme it recognized and was wrong at 10/10 (disaster girl, 2004 vs 2005). gemini said 7/10 on every meme.
• qwen's thinking loops: 66s on charlie re-reading one line ~40 times, ~70% of the obama reasoning stuck on the year.

conclusion: gemini is more accurate and ~10x faster on short memes. qwen reasons at length, describes the long video in more detail – 21 topics to gemini's 19 and reads the lower-third caption, gets mandarin right, gets mandarin right where gemini misheard, and is ~2.5x cheaper at list price. the audio gap is qwen studio's, not qwen's.

watch the full test via link: