Gemini Omni Flash 1.2 Lands at #2 on Video Editing — and the Error Bars Are the Story

Artificial Analysis put G​emini Omni Flash 1.2 on three video leaderboards on 10 November, six days after G​oogle made the model generally available. The headline writes itself: #2 on AA-Video-Editing v1.1, ahead of the #5 held by G​emini Omni 1.1 Flash, with a climb from #12 to #6 on AA-Video-T2V v2.1 and from #17 to #14 on AA-Video-I2V v2.1. The useful reading is in the margins. Two of those three moves sit inside or against overlapping 95% intervals, one of the three was measured on a reduced prompt set, and the cleanest separation of the lot belongs to the smallest rank change in the set.

Three boards, three different answers about who is best

The first thing the new rows establish is that there is no single video ranking to win. Wan 3.0 leads the editing board at 1,192 and the text-to-video board at 1,175, and then lands 11th on image-to-video at 1,108. Dreamina Seedance 2.5 tops image-to-video at 1,161 while sitting fourth on editing and fourth on text-to-video. A model's position is a property of the task, not of the model, and any pitch that quotes one number is quoting one task.

Against that backdrop, G​emini Omni Flash 1.2's profile is unusually flat. It is second on editing, sixth on text-to-video and fourteenth on image-to-video — competitive everywhere, dominant nowhere. For a model whose headline feature is conversational editing, that flatness is arguably the point: it is the only row in the top ten of the editing board that also appears in the top ten of text-to-video.

Twenty-six Elo, and two intervals that nearly touch

The move from #12 to #6 on AA-Video-T2V v2.1 is the one that will get screenshotted. It is a 26-point gain, 1,128 to 1,154. The published intervals are 1,115 to 1,141 for G​emini Omni 1.1 Flash and 1,140 to 1,168 for 1.2. They overlap by exactly one point.

That is not a reason to dismiss the gain. It is a reason to stop treating a rank as a measurement. Artificial Analysis derives these intervals from blind pairwise votes, and at 5,056 samples for 1.2 against 7,943 for 1.1 the newer model has the thinner evidence base because it has been on the board for a day. A 26-point gap that touches at the boundary is best read as "probably better, not yet separable at this sample size" — which is a different claim from "moved up six places", even though both describe the same two rows.

The editing board tells a similar story with a smaller gap: 1,164 against 1,146, an 18-point difference whose intervals overlap across 1,152 to 1,156. The image-to-video board is the opposite case and the more interesting one. There the gap is 42 points, 1,086 against 1,044, and the intervals do not touch at all — 1,071 to 1,101 against 1,030 to 1,058. The rank moved three places instead of six, and it is the only one of the three comparisons that survives a strict reading of its own error bars.

Sort the three accordingly and the ordering inverts: the smallest rank change is the strongest evidence, the largest rank change is the weakest. Rank arithmetic hides that, because ranks are computed from point estimates with the intervals thrown away.

The editing score comes with a footnote attached

G​emini Omni Flash 1.2's #2 on AA-Video-Editing v1.1 carries a note on the row itself: the model "declined some of our test set through its copyright and safety filters", and its result reflects the remaining prompts. Every row is also flagged provisional.

Read that plainly. The score is an average over the prompts the model was willing to answer, and the prompts it refused are exactly the ones where an editing model is asked to alter real footage, reproduce a logo, or work on a recognisable subject — the tasks that separate a demo from a delivery. A model that declines the hard third of a test set can outscore a model that attempts all of it without being better at the job. Nothing in the published result lets you correct for that, and the direction of the bias is upward, not downward.

This is not an argument that the number is wrong. It is an argument that #2 on the editing board is a claim about quality on permitted prompts, and that the permitted set is smaller for this model than for the ones around it. If your work involves archival footage, brand assets or likeness, the operator's own filter behaviour will decide more of your outcomes than an 18-point Elo gap will, and the leaderboard will never show you that because the refusals are not published.

$6.26 a minute is G​oogle's own price list, arrived at independently

The pricing column on all three boards lists G​emini Omni Flash 1.2 at $6.26 per minute of generated video. Artificial Analysis defines that column as the cost to generate one minute on the creator's API at the settings it benchmarks, 1080p where offered — a normalised figure, not a quote.

Here it happens to reconcile exactly with the vendor's own arithmetic. G​oogle bills video output at 8,688 tokens per second for 1080p with upsampling, at $12.00 per million tokens. Sixty seconds is 521,280 tokens, which is $6.255. The board's $9.12 for G​emini Omni 1.1 Flash reconciles the same way against $17.50 per million. Two independent parties arriving at the same cent is worth more than either number alone: it means the 30% cut G​oogle announced on 5 November is visible in a third party's cost model, not just in a press release.

What the column does not carry is the resolution you would actually pick. At 720p the same model bills about $0.07 per second, roughly $4.20 a minute; at 360p, $0.023 per second, about $1.38. A team drafting at 360p and finishing at 1080p pays a blended rate that appears nowhere on the board, and the board's single number will overstate the drafting half of that pipeline by a factor of four.

What a preference arena cannot tell you

These are blind human-preference votes, not a benchmark suite. Nobody is measuring whether a fifteen-second clip holds a character's jacket colour across a cut, whether the audio stays in sync at the fourth extension turn, or whether the model respects a six-reference-image brief. Those are the properties G​oogle's 5 November changelog actually claims — improved character and subject consistency across up to six reference images, clearer lip-sync, distinct multi-speaker conditioning — and none of them is what a pairwise "which clip is better" vote measures.

The votes also cannot see the failure modes that matter operationally. A model that declines a prompt scores nothing; a model that returns a technically valid file with drifting faces scores well on a single-clip comparison and badly over a twenty-shot sequence. Arena Elo is a per-clip instrument being used to answer a per-project question.

Where this leaves a pipeline, and what we route

G​emini Omni Flash 1.2 is not in OrcaRouter's catalogue. We checked the live model list and there is no Omni entry of any version, so if the model is what you need, G​oogle's own API is the path and we would rather say that than imply otherwise. What the routing layer changes is the part of a video pipeline that is not video: the shot list, the prompt expansion, the captioning pass and the post-hoc audit are all text work, and those run on models we do carry — including on one key, at provider list price with 0% markup, so a vendor price move reaches the bill the day it happens rather than at the next contract.

There is a concrete version of that argument in this release. The 1.2 video rate is 30% below 1.1 and the 1.1 endpoint stops being separately billable on 3 December, when G​oogle starts routing it to 1.2 automatically. A pipeline that hard-coded the old per-second constant now has three weeks to notice. A pipeline that reads the rate from a gateway that passes list price through does not have the problem, because there is nothing to re-hardcode.

What would change this read

Three things, in order of usefulness. A published refusal rate for the editing board would turn the #2 into something comparable; today the note tells you the test set shrank but not by how much. A second week of votes on 1.2 would narrow the text-to-video interval enough to settle whether 26 Elo is a real gap or a sampling artefact — the current overlap is one point wide, which is the least informative place for an interval to sit. And any independent measurement of the reference-image consistency claim would be worth more than all three boards combined, because that claim is the one G​oogle is actually selling and the one no arena is testing.

Until then the honest summary is narrow. G​emini Omni Flash 1.2 is a better video model than G​emini Omni 1.1 Flash on all three boards, clearly so on image-to-video and probably so on the other two, at 30% less per token, with a safety filter that removes part of its editing test set and an operator who will route the old endpoint to the new one in three weeks whether or not you have tested it.