Gemini Omni Flash 1.2 vs Qwen 3.8 Omni Flash: One Makes the Footage, the Other Reads It

The question "which is better, G​emini Omni Flash 1.2 or Qwen3.8-Omni-Flash" has a short answer that most comparison pages will not give you: they cannot be compared, because one of them has never produced a frame of video in its life. G​emini Omni Flash 1.2 is a video generator with native audio that takes text, images and reference video in and returns an MP4 out. Qwen3.8-Omni-Flash is a multimodal reader that takes text, image, audio and video in and returns text out — no video, no image, no audio generation at all. They are two different machines that a video team ends up buying in the same quarter. The decision in front of you is not A or B. It is how much of each, and the arithmetic that answers it is duration, not quality.

The output column settles it before the benchmarks do

Every comparison of these two models starts with a leaderboard that does not exist. Artificial Analysis runs three video boards — editing, text-to-video and image-to-video — and G​emini Omni Flash 1.2 is scored on all three, currently second, sixth and fourteenth. Qwen3.8-Omni-Flash is on none of them. Not ranked low: absent, because there is no video output to rank. A head-to-head Elo between these two is not a number we are withholding. It is a number nobody can compute.

What you can compare is the interface. G​emini Omni Flash 1.2 returns video. Qwen3.8-Omni-Flash returns tokens. Everything downstream of that single difference follows mechanically.

• Input — G​emini Omni Flash 1.2 takes text, up to six reference images, and reference video clips; Qwen3.8-Omni-Flash takes text, images, audio and video, including footage over an hour long in one call.

• Output — G​emini Omni Flash 1.2 returns video with native synchronised audio; Qwen3.8-Omni-Flash returns text only.

• Context — G​emini Omni Flash 1.2 accepts 131,072 input tokens and returns 65,536; Qwen3.8-Omni-Flash accepts just under a million input tokens (991,808) and returns up to 131,072.

• Duration ceiling — G​emini Omni Flash 1.2 generates three to fifteen seconds per call, extendable turn by turn to 30 seconds; Qwen3.8-Omni-Flash watches by the hour, and its ceiling is a token budget rather than a runtime.

• Price — G​emini Omni Flash 1.2 bills video output at $12.00 per million tokens with text and media input at $1.50; Qwen3.8-Omni-Flash bills $0.15 per million input and $0.47 per million output.

What fifteen seconds of generated video costs, against an hour of watched video

Put the two price lists on the same axis and the asymmetry stops being abstract. G​oogle bills generated video at a fixed token rate per second of footage: 5,792 tokens per second at 720p, 8,688 at 1080p with upsampling. At $12.00 per million, a fifteen-second 720p clip is roughly $1.05 in output tokens, and a full minute of 1080p is $6.26 — the figure Artificial Analysis prints in its cost column, which reconciles to the cent with G​oogle's own rate card.

Now the other side. Qwen3.8-Omni-Flash does not bill by the second of footage; it bills by the token, and video frames are tokenised at a rate set by the sampling configuration. An hour of footage at a modest one frame per second arrives as tens of thousands of visual tokens, not hundreds of thousands, and at $0.15 per million input the ingest of that hour lands in cents rather than dollars. The exact figure depends on the frame rate and resolution you choose, which is the point: watching is cheap enough that you tune it for accuracy, while generating is expensive enough that you tune it for length.

That asymmetry is the whole purchasing decision. A team that budgets these two as substitutes — "we have X dollars for video AI" — will systematically under-buy the cheap side and over-constrain the expensive one. A team that budgets them as a pipeline buys an hour of machine review for less than one second of generated footage, and then spends the generation budget only on the clips that survive.

The workflow nobody writes about: what happens before the prompt

Here is where the pairing stops being theoretical. The expensive failure in a video pipeline is not a bad render. It is a good render of the wrong thing, discovered after four extension turns and six reference images have been billed.

A reader model changes that sequence. Feed the source footage, the client's existing reel, the product walkthrough, the fifty-minute interview — whatever the new clip has to match — into Qwen3.8-Omni-Flash and ask for the things a prompt needs: the shot list with timings, the dominant colour and lighting per scene, the on-screen text, the speakers and when each one talks, the cuts that must not be crossed. That output is text, and text is what you feed a generator. G​emini Omni Flash 1.2 does not accept a video as an instruction set; it accepts a prompt and up to six reference images, and it cannot watch an hour of anything to work out what you meant.

The reverse direction is weaker but real. Once a clip exists, the same reader model can audit it — does the jacket colour hold across the cut, is the audio in sync at the fourth extension, does the on-screen label survive to the last frame. Those are the exact properties G​oogle's 5 November changelog claims for 1.2 and the exact properties no preference arena measures, because a pairwise "which clip looks better" vote cannot see a continuity error that only appears at second twelve.

The honest limit: none of this is vendor-blessed workflow. A​libaba's documentation describes Qwen3.8-Omni-Flash as a general multimodal model, not as a video-production tool, and G​oogle's guide describes no ingestion path for its competitor's output. The pattern is one practitioners assemble, and it works because both endpoints speak plain text at the seam.

Where each one is the wrong tool

Qwen3.8-Omni-Flash is the wrong tool the moment the deliverable is a file with a picture in it. There is no image mode, no video mode, no audio mode. If your brief says "make me a fifteen-second cut", the model can tell you what should be in it and cannot make it.

G​emini Omni Flash 1.2 is the wrong tool for anything longer than half a minute, and for a specific set of markets. Three- to fifteen-second generation with extension capped at 30 seconds means a two-minute explainer is a stitching job, and every stitch is a place where continuity can break — the thing a reference-image budget of six is meant to hold together and does not guarantee. Editing and extension of uploaded video are unavailable in the EEA, Switzerland and the UK, so the conversational-editing pitch does not exist for users there. There is no free tier on either version, and no system instructions, temperature, top_p or negative prompts — the model takes a prompt and reference media and nothing else. Every output carries a SynthID watermark.

And one asymmetry that matters commercially: 1.1's endpoint is routed to 1.2 automatically on 3 December 2026, so the older G​emini version is not a long-term fallback. Qwen3.8-Omni-Flash has no such forced migration attached to it. If you are building a pipeline that needs a stable reader, the reader is the stable half of this pair.

Where the routing layer actually sits

Neither of these two endpoints is in OrcaRouter's catalogue, and we would rather state that than let a comparison page imply otherwise. G​emini Omni Flash 1.2 goes through G​oogle's API; Qwen3.8-Omni-Flash goes through A​libaba Cloud Model Studio. What we do carry is the layer on either side of them.

The ingest-and-audit half of the pipeline above is text and vision work, and that is the half that runs on models we route — including Qwen3.8-Flash, the text-only sibling in the same family, and the Qwen3-VL instruct models for frame-level description, all reachable on one key at provider list price with 0% markup and automatic failover if an upstream degrades. A routing rule can send a fraction of audit requests to a second model without a deploy, which is how you find out whether the cheap reader is good enough before you commit a production path to it. The generation call stays where it has to stay: direct to G​oogle, on G​oogle's meter, with G​oogle's rate change landing on your invoice on G​oogle's schedule.

That split is worth naming plainly because it is the honest shape of a 2026 video stack. No single vendor sells both halves well, and the teams that suffer are the ones that tried to make one contract cover a reader and a renderer.

How to decide, in the order the costs arrive

Start with the reader, because it is cheap and it produces the artifact everything else consumes. If you cannot get a usable shot list, colour report and transcript out of your source footage, no amount of generation quality will save the project — you will be prompting blind. Qwen3.8-Omni-Flash is a reasonable first reader at $0.15 per million input tokens, and its near-million-token context means an hour of footage fits without a chunking strategy.

Then measure your actual clip lengths. If your median deliverable is under fifteen seconds — social cuts, product inserts, B-roll beats — G​emini Omni Flash 1.2 covers it in a single call at about $0.07 per second at 720p, and the six-reference-image budget is enough to hold a subject across that span. If your median is over thirty seconds, you are buying a stitching pipeline whatever the per-token price says, and the 30-second extension cap is the constraint that governs your schedule rather than your bill.

Then price the failure, not the success. A refused editing request on 1.2 costs nothing and returns no file; a clip that drifts at second twelve costs the whole render plus the reviewer's time. The second failure is the one a reader model catches, which is why the cheap half of this pair tends to pay for itself first.

What you should not do is treat the two as a bake-off. There is no shared leaderboard, no common output, and no scenario in which one replaces the other. The useful comparison is a ratio — how many minutes of footage you can afford to have watched for every second you can afford to generate — and for most teams in late 2026 that ratio is measured in the hundreds.