Gemini Omni 1.2 Flash vs Qwen3.8-Omni-Flash: The Word "Omni" Is Doing Two Jobs

Alibaba's own published estimate puts the cost of reading an hour of 720p video with Qwen3.8-Omni-Flash at about twenty cents. Google's list price for an hour of generated 720p from Gemini Omni 1.2 Flash is $252. Both models are called Omni. Both accept text, image, audio and video. One reads footage and answers in text; the other writes footage and answers with a file. The price gap between them is not a competitive signal, because they are not competing — it is the cost of manufacturing a second of reality versus the cost of describing one.

This matters because the comparison is being made constantly, and almost always with the wrong model on the Google side. Every "Qwen Omni versus Gemini" article in circulation — including the benchmark table Alibaba itself published — measures Qwen3.8-Omni-Flash against Gemini 3.8 Flash, a text-out model at $0.75 per million input tokens. That is the correct rival. Gemini Omni 1.2 Flash, generally available since 5 November 2026, is a different product on a different meter, and there is no published head-to-head between the two anywhere.

What Qwen3.8-Omni-Flash actually is

Alibaba's Qwen team shipped it on 18 September 2026 as the successor to Qwen3.5-Omni-Plus, built on the Qwen3.8-Flash-Next architecture. The first thing to get straight is the output modality, because the name invites the wrong assumption: it produces text and only text. It does not generate video, and unlike its predecessor it does not generate speech either. Everything omni about it is on the input side.

On that side it is unusually capable. The context window is one million tokens, with roughly 991,000 available for input and 131,000 for output. A single call can take an hour of continuous video. It handles dual-channel stereo and four-channel spatial audio, which is how Qwen gets its headline party trick — locating a sound source in a room and pointing at it, the kind of capability a robotics or meeting-analytics team has been waiting for. Speech recognition covers 74 languages plus 39 Chinese dialects; text-to-speech covers 29 languages in the sibling models. There is a realtime variant, Qwen3.8-Omni-Flash-Realtime, that runs over WebSocket or WebRTC with speech output and 56 voices on a 192K context.

Pricing is the aggressive part: $0.15 per million input tokens, $0.47 per million output, and $0.016 per million for cached input. Weights are closed. It is API-only, through Alibaba Cloud Model Studio and Qwen's own endpoints, OpenAI-compatible, in six regions.

What Gemini Omni 1.2 Flash actually is

Google's model goes the other way. Text, images, audio and video go in; three to fifteen seconds of video with synchronized audio comes out, in one-second increments, at 16:9 or 9:16. You can draft at 360p or 720p and upscale to 1080p or 4K. A request can carry up to six reference images, a motion clip of up to three seconds, and audio cues, and multi-turn editing lets you swap a product, change the weather or reframe the shot without regenerating the whole thing. Extension adds three to ten seconds per turn up to a 30-second ceiling. Everything is watermarked with SynthID.

The 1.2 release is a price cut as much as a capability release: video output fell from $17.50 to $12.00 per million tokens, roughly $0.07 per second of 720p, a 30% drop. Text input stays at $1.50 per million and text output at $9.00 per million. Paid-tier rate limits doubled the same day. Google published no benchmark scores with the release — only qualitative claims about reference adherence, lip-sync and camera continuity.

Three cost traps that only show up when you run both

The reasoning budget is billed as output. Qwen3.8-Omni-Flash defaults its thinking mode to "xhigh", with a reasoning budget of up to 262,000 tokens that is charged at the output rate. A single request that thinks hard about a long video can cost around $0.123 in reasoning alone — which is more than six hundred times what the video-reading estimate suggests, and the reason some teams report the model as expensive when their actual footage volume is trivial. Cap the thinking budget before you benchmark anything.

The per-hour video figure is a vendor estimate. Alibaba does not publish a tokens-per-second rate for video ingestion, so the "twenty cents per hour" number cannot be independently derived from the price sheet. The $0.15/$0.47 rates are verifiable; the per-hour conversion is Qwen's arithmetic, and one published analysis has already flagged the 98%-cheaper-than-Gemini claim as unverifiable for exactly this reason.

The render cost multiplies by takes, not by deliverables. A fifteen-second clip is $1.05. Six cutdowns is $6.30 — but nobody gets six out of six on the first pass. At three attempts per keeper, the same six cutdowns cost $18.90, and that is the number that goes in the budget. Google also caps a stitched sequence at 30 seconds, so anything longer is a second generation pass, not an extension.

A worked pipeline: thirty minutes in, six clips out

Here is where both models earn a place, and where the cost distribution is counterintuitive.

Read the master. Qwen3.8-Omni-Flash takes the 30-minute source video and returns a transcript with speaker turns, on-screen text, a shot list with timecodes, and a short list of the moments where someone says something quotable. At Alibaba's estimate that is roughly ten cents of input. Add a thinking pass to pick the six best moments and you are still under a quarter — as long as the budget is capped.

Render the cutdowns. Six fifteen-second clips from Gemini Omni 1.2 Flash, three takes each: $18.90. Reference images of the product keep the packaging consistent across all six; audio cues keep the dialogue intelligible; multi-turn editing fixes the two that drift.

Read the outputs back. The six finished clips go back into Qwen3.8-Omni-Flash to confirm the on-screen text is legible, the dialogue matches the script, and nobody's lip-sync slipped a frame. Six clips is ninety seconds of video. The reading cost is negligible; the checking is what makes the render bill defensible.

The pattern is general: on this pair, analysis is cheap and generation is expensive, by two to three orders of magnitude. Teams that budget for the reverse — assuming the "intelligent" model must be the costly one — under-provision the render line and over-provision the read line every time.

The benchmark record, and whose numbers these are

Everything published about Qwen3.8-Omni-Flash comes from one place: Alibaba's own evaluation, released as an arXiv paper on 22 September 2026. Against Gemini 3.8 Flash it wins WildClawBench-MM 71.0 to 58.9, SpotSoundBench 67.2 to 39.7, and meeting diarization by an enormous margin — a diarization error rate of 3.35 against 72.6. It loses OmniVideoBench 63.4 to 65.2 and AgenticVBench 36.8 to 45.0. None of this has been independently reproduced, and none of it involves Gemini Omni 1.2 Flash, which does not read video into a context window at all.

Google's side of the ledger is emptier. The Gemini Omni 1.2 Flash announcement contains no numbers — no Elo, no win rate, no benchmark suite. The comparisons are qualitative: better reference matching than 1.1, more fluid dialogue, clearer lip-sync, fewer jump cuts. NVIDIA's and Alibaba's releases at least arrive with tables; Google's arrives with adjectives.

So the honest position on "which is better" is that the question has no measured answer and will not have one until someone runs both on the same task. What can be said with confidence is narrower and more useful: Qwen's numbers were earned against a text-out model at a fifth of the input price, and they will not transfer to a video generator, because the two systems are scored on completely different instruments.

Where the two vendors genuinely overlap

Not on the model. On the enterprise contract. Both are closed, both are API-only, both are region-locked in ways that matter — Alibaba serves Beijing, Singapore, Hong Kong, Tokyo, Frankfurt and Virginia; Google's editing features for uploaded video are unavailable in parts of Europe. A media company choosing a vendor is choosing a data-processing agreement and a jurisdiction as much as a capability.

It is also worth noting what neither will do. Qwen3.8-Omni-Flash cannot produce a frame, and Google's Omni 1.2 Flash cannot answer a question about a ten-minute clip. Anyone consolidating onto one vendor to cut procurement overhead will find that the consolidation costs them a capability, whichever direction they consolidate. The migration note from Google makes the same point from the other side: from 3 December 2026, requests to gemini-omni-1.1-flash are answered by 1.2 at 1.2 prices, with no code change — a reminder that closed endpoints move on the vendor's schedule, not yours.

What OrcaRouter covers, and what it does not

Neither of these two models is on OrcaRouter, and this piece will not pretend otherwise. What is on the platform is the adjacent capacity to build the pipeline above without a second and third contract.

The read side is covered by Qwen3.8 Flash: text, image and video in, text out, a one-million-token context, long-video analysis, billed at $0.15 per million input and $0.47 per million output — the same price and the same profile as the Omni variant, minus the spatial-audio features. The render side is covered by Kling 3.0 Omni at $0.084 per request for text-to-video and image-to-video with 3–15 second clips. Both sit behind one key, which means a fan-out that sends the same source to a reader and a writer is a routing rule rather than an integration project; automatic failover means a provider wobble does not stop a campaign mid-render; and the 0% markup is a straight pass-through of provider list price, so the day a vendor moves a rate the number on our side moves with it.

How to decide in the next ten minutes

If your problem is that you have footage and no way to search it, subtitle it, or find the moment where the presenter mentions the price — buy the Qwen side. Cap the thinking budget first, and measure your real ingest rate rather than trusting the per-hour estimate.

If your problem is that you have a product shot and a deadline and no camera crew, buy the Google side, and budget three takes per keeper rather than one.

If you have both problems, which is what most teams actually have, stop looking for a winner. The two models named in the first paragraph of this article will never appear in the same benchmark table, because one of them produces a file and the other produces a paragraph about the file. The only number that connects them is the one nobody publishes: how many hours of video your organisation reads for every hour it makes.