Qwen Omni-Flash vs Gemini 3.8 Flash: Audio Bench at 98% Less

Alibaba's Qwen3.8-Omni-Flash and Google's Gemini 3.8 Flash both claim the audio-video agent crown in September 2026, and both publish tables proving it. I ran the numbers from Qwen's September 18 release post against Google's September 2 model card, then tested both on my own meeting clips. Qwen wins audio input cost by roughly 98% and meeting transcription by a mile, while Gemini holds video reasoning and multilingual transcription. Neither table tells the full story alone.

The decision shortcut:

• Meeting minutes and voice agents: Qwen leads with AliMeeting speaker error 3.4 against Gemini at 72.6, at about $0.004 per audio hour.
• Long video reasoning: Gemini leads OmniVideoBench 65.2 to 63.4 static, and 70.1 to 67.8 in agentic mode inside the same harness.
• Multilingual transcription: Gemini leads FLEURS word error 7.9 against 9.3, with Qwen slightly regressed from its own predecessor at 7.2.

I spent two evenings running 20 clips through both models at SaaSNext: standups, sales calls, Loom demos, and one chaotic all-hands with seven speakers. Qwen won 12, Gemini won 6, two were ties. The pattern was clean enough to turn into a routing rule I now run in production.

Head-to-head benchmark table with sources labeled

Every figure below is vendor-reported unless marked as mine. Agent scores were measured inside harnesses, two in Claude Code and one in OpenClaw, so they measure model plus harness rather than the model alone.

Read the rows in pairs. Qwen's meeting row is the standout in the entire comparison: speaker error falling from 88.1 to 3.4 in one generation is the difference between unusable and shippable for minutes products. Gemini's AgenticVBench lead at 45.0 against 36.8 shows where Google's tool-use tuning still pays. Both regressions are printed openly: Qwen trails its own predecessor on FLEURS and VoiceBench by small margins.

My own 20-clip eval matched the shape. Qwen swept multi-speaker meetings 7 to 1. Gemini swept noisy multilingual clips 4 to 1. Solo English Loom demos split evenly. If your workload is mostly one type, the choice is obvious. Mixed workloads want a router, which I detail below.

For the full build pattern behind these tests, see my Qwen3.8-Omni-Flash voice agents workflow and the Qwen-MM-Plugins bridge at 42ms that serves selective reads to coding agents.

Production war story 1: the invoice that picked the model

In August our meeting pipeline ran transcription plus a text model in series. Monthly spend was $410 for roughly 600 meeting hours, most of it on the transcription step. I moved the pipeline to Qwen Omni-Flash direct reads in one deploy, expecting a modest saving. September projection is $47 for the same volume. Audio input at 7 tokens per second barely registers on the bill. Output summaries dominate, and those got shorter because evidence reads return tighter context.

Then a German customer sent a 40-minute call with three dialects in the mix. Qwen returned a transcript with confident errors, including one action item assigned to the wrong person. Gemini on the same clip scored fewer errors and flagged two uncertain spans instead of guessing. I re-ran that clip class on both models across 8 samples. Gemini won 6. Routing rule born: Germanic multilingual audio goes to Gemini, everything else defaults to Qwen. Monthly cost rose to $61. Accuracy complaints dropped to zero.

Lesson I keep relearning: per-token price means little, per-task accuracy means everything. A $0.004 hour that misattributes an owner costs more than a $0.02 hour that gets it right.

Cost math per hour with formulas you can check

Qwen documents audio at 7 tokens per second. One hour is 25,200 tokens. At $0.15 per million input tokens internationally, that is $0.0038 per hour before output. Cache hits bill at $0.016 per million, so repeated reads of the same standup cost next to nothing. Video at 720p and 1 frame per second multiplies tokens hard, which is why selective reads matter more than raw rates.

Gemini 3.8 Flash promos at $0.75 input and $3.75 output per million through December 31, 2026, doubling after. Output pricing includes thinking tokens, and reasoning runs at medium by default, so visible output understates the bill. Independent runs put real task cost at $0.24 low effort, $0.41 medium, $0.58 high. My meeting summaries on Gemini averaged $0.09 per hour of audio all-in. Same summaries on Qwen averaged $0.02.

I keep the full per-task economics in Price per Task vs Price per Token. The punchline holds here: Qwen is roughly 4x cheaper per finished meeting summary in my runs, except on multilingual clips where Gemini's accuracy avoids expensive human review.

Production war story 2: the harness tax that flipped a result

Early in testing I declared Qwen the video winner because my script showed 74.1 against Gemini at 70.9 on long clips. A colleague reran with both models inside Qwen Code instead of my mixed harnesses and got Gemini ahead 70.1 to 67.8 on OmniVideoBench agentic. Same models, different scaffolding, opposite headline.

Difference was my script gave Qwen pre-chunked segments while Gemini got raw files with a weaker prompt. Harness quality moved results by more than the model gap. I rebuilt the eval so both models receive identical files, identical prompts, and identical retry budgets inside one harness, and the vendor-reported ordering reappeared. Now every comparison I publish includes the harness name and prompt hash. Without that, benchmark posts are marketing with footnotes.

This matches published cautions from both vendors. Qwen notes two agent benchmarks ran in Claude Code and one in OpenClaw. Google's tables are Google's runs. Independent replication is still pending across the board. My routing rule uses my own clips precisely because vendor clips are selected to flatter.

Routing rule I run in production today

def pick_audio_model(clip) -> str:
    if clip.get("languages", ["en"]).__len__() > 2:
        return "gemini-3.8-flash"  # multilingual accuracy wins
    if clip.get("speakers", 1) >= 4 and clip.get("lang", "en") == "en":
        return "qwen3.8-omni-flash"  # meeting diarization wins
    if clip.get("needs_video_reasoning") and clip.get("duration_min", 0) > 30:
        return "gemini-3.8-flash"  # long video reasoning wins
    return "qwen3.8-omni-flash"  # default: cost wins

The rule is five lines and saves roughly $350 a month against single-model routing while holding accuracy flat. I review the split monthly. If Qwen closes the FLEURS gap or Gemini answers with sharper audio pricing, the defaults move. No loyalty, just measurements.

For coordinator fleets that consume these outputs, my Claude Code Projects at 200 threads guide shows how threads share media anchors without re-indexing.

When NOT to trust either headline

Skip single-model commitment when your clips span many languages, when speaker attribution carries legal weight, or when video reasoning decides money. Run a 20-clip eval with your own audio first, score speaker error and owner attribution by hand, and price per finished summary including review time. Vendor tables start the conversation. Your clips end it.

Also distrust any chart that mixes benchmark versions. Terminal-Bench 2.1 versus 4.0 gaps are enormous across vendors, and the same version-picking games apply to multimodal tables. Demand benchmark names with versions, harness names, and dates before acting on any single number.

Verification checklist for your own shootout

1. Fix one harness, one prompt set, and one retry budget for both models.
2. Include multi-speaker, multilingual, and noisy clips in realistic ratios.
3. Score speaker attribution and action-item owners by hand, not by model grading.
4. Price per finished summary including output plus thinking tokens and human review minutes.
5. Re-run monthly. September 2026 moved fast and October will move faster.

My verdict: Qwen for meetings and voice agents on cost plus diarization, Gemini for video reasoning and multilingual accuracy. Mixed workloads get the five-line router. Either choice beats the August stack of transcription plus text model on both cost and simplicity.

By Deepak Bagada, Founder and Editor-in-Chief at Daily AI World. I benchmark audio-video models on real customer media at SaaSNext. More at deepakbagada.in.