Gemini Omni Flash 1.2 vs Gemini Omni 1.1 Flash: Two Leaderboards Say Yes, One Says Not Yet

On 3 December 2026, every request sent to G​emini Omni 1.1 Flash starts landing on G​emini Omni Flash 1.2 instead, billed at the new rates, with no change to your payload. G​oogle published that routing date in the same changelog entry that made 1.2 generally available on 5 November. So the version question is not whether you migrate. It is whether you finish measuring before the switch happens to you, and the independent evidence for doing that is more mixed than the price cut suggests.

What the same week of voting actually showed

Artificial Analysis ran both versions on all three of its video boards and published the pair within days of each other. That is the rare case where a version comparison does not have to be assembled from two different measurement campaigns at two different times.

On AA-Video-I2V v2.1 the case is closed. G​emini Omni Flash 1.2 sits at 1,086 with a 95% interval of 1,071 to 1,101; G​emini Omni 1.1 Flash sits at 1,044 with 1,030 to 1,058. Forty-two points, no overlap, 14th against 18th. Whatever else is uncertain about the new model, its image-to-video output is measurably preferred to its predecessor's by a margin the sample supports.

On AA-Video-T2V v2.1 the case is open. The 26-point gap — 1,154 against 1,128 — comes with intervals of 1,140 to 1,168 and 1,115 to 1,141. They share a single point. On AA-Video-Editing v1.1 the gap is 18 points, 1,164 against 1,146, and the intervals overlap across four. Both boards rank 1.2 above 1.1. Neither separates them.

That pattern is worth more than any one number. A new checkpoint that wins clearly on one task, probably on a second and marginally on a third is the ordinary shape of a point release. It is not the shape of the generational jump the 15-second clips and doubled rate limits might suggest.

The price is the part with no error bars

Video output falls from $17.50 to $12.00 per million tokens, a 30% cut, and it is the only claim in the release that needs no confidence interval. G​oogle bills at a fixed token rate per second of footage, so the arithmetic is not a matter of preference: at 720p, 5,792 tokens per second, the cost goes from about $0.10 to about $0.07 per second. At 1080p with upsampling it is about $0.104 per second, and at 4K about $0.209. Artificial Analysis normalises both versions to $9.12 and $6.26 per minute of 1080p, which is the same 30% seen from the other end of the same meter.

Input stays at $1.50 per million tokens for text, image and video, and text output stays at $9.00. Nothing moved in the batch, flex or priority columns because 1.2 has none of them — standard tier only, and no free tier on either version.

Do the clip arithmetic before the routing date rather than after. Fifteen seconds at 720p on 1.2 is roughly $1.05 in output tokens. The same fifteen seconds could not be produced on 1.1 in a single call, since 1.1 caps at ten seconds; the honest comparison is ten seconds for about $1.00. The cut is real, but a large share of it is spent buying the extra five seconds, and a team that never needed longer clips will see a smaller invoice change than the percentage implies.

Where the fifteen seconds stops being free

The duration bump comes with edges that are documented rather than discovered, and three of them change what a request is allowed to contain.

• Interpolation — first-and-last-frame control works from three to ten seconds on both versions; asking for eleven seconds or more with two reference frames returns an invalid argument error on 1.2.

• Reference video — three-to-ten second outputs accept three reference videos of up to three seconds each, while eleven-to-fifteen second outputs accept one. Lengthening a shot can silently invalidate a request that passed with three.

• Extension — extension runs three to ten seconds per turn with a ten-second lookback, capped at 30 seconds total, and only at the end of the clip. On 1.1 the editing and extension ceiling was ten seconds of input.

• Reference images — up to six per request, across all three modes, which is the change G​oogle leads with in the launch language about character and subject consistency.

• Region — editing and extension of uploaded video is unavailable in the EEA, Switzerland and the UK, a restriction carried in the guide. If your users are in those markets, the conversational-editing pitch does not apply to them at all.

• Limits that did not move — no system instructions, no temperature, no top_p, no negative prompts, no provisioned throughput, no YouTube sources, no multi-video reasoning. Every output carries a SynthID watermark.

A "Text" output row that does not mean what it looks like

The model card lists 1.2's output as "Video with native Audio, Text" against 1.1's plain "Video", and that difference has been read as a new text channel. It is not one. The Omni guide describes no text-output mode for 1.2; text appears rendered inside the frame — labels, signage, axes — and the request shape is still a video response. REST responses carry a steps array whose "thought" entry is a reasoning trace, not an answer. If a pipeline needs a caption string back from the API, neither version returns one, and the difference in that row is a description of what the model draws, not of what it gives you.

The audio row is a real change but a change in quality, not in kind. Both versions generate native synchronised audio. G​oogle's launch claims more fluid dialogue, clearer lip-sync and distinct multi-speaker conditioning, and says references can carry audio cues and one-second voice samples — those reference paths are allowlisted on G​oogle Cloud, audio input is unsupported on the model itself in this version, reference-video audio is ignored, and voice editing is not available. You can prompt for a voice; you cannot edit one afterwards. None of those quality claims has an independent test behind it, and the three leaderboards above measure visual preference, not whether the words match the mouths.

The migration that breaks downstream, not the request

Because the payload does not change, the failure surface is everything the request feeds. Four places to check before 3 December, in the order they tend to break:

Duration defaults. A request that omitted duration was capped at ten seconds and still defaults to ten, so untouched calls behave identically. What breaks is validation that treated ten as the maximum.

Frame-pair conditioning. Any code path that pairs interpolation with a duration above ten now errors instead of truncating quietly — better behaviour, louder failure.

Cost models. Every per-second constant derived from the 1.1 rate is 30% too high at 720p. If you bill internally per render, this is the line to change while the old prices are still correct, because after the switch the number will be wrong in the direction that flatters your margin.

Filter behaviour. The editing board's row for 1.2 carries a note that the model declined part of the test set through copyright and safety filters, and the published score reflects the surviving prompts. A production pipeline that assumed every request returns a file needs a refusal path on the new version specifically. The note appears only on 1.2's row, not on 1.1's.

Two ways to make the switch, and which one fits

If your content is measured and you like the output, migrate early and use the three weeks to re-measure on your own footage. The price cut and the duration bump are worth having, the image-to-video gap is real, and the alternative is being moved on 3 December with no data either way.

If your content is high-volume and you have not looked at it since the last change, the useful posture is to run both. Keep 1.1 on the production path while it is still separately addressable, shadow a percentage of traffic to 1.2, and compare on the axis the boards do not measure — how often a clip survives to delivery, not how often it wins a pairwise vote. That is a routing problem rather than a model problem, which is the layer we operate. Neither Omni endpoint is in OrcaRouter's catalogue today, so this is not a pitch for us to serve them; it is the reason a version bump hurts most in the pipelines that have exactly one path and no way to run a second beside it. Where we do earn the placement is the surrounding text work: the prompt-expansion and shot-list calls that feed a render can be composed against several models through one endpoint at provider list price, with automatic failover if one degrades, and a routing rule can send a fraction of requests to a different model without a deploy.

The one thing not to do is pin 1.1 past 3 December and discover the change from an invoice. The endpoint routes regardless of what your config says.

The version comparison, stated as a version comparison

1.2 is cheaper per second, longer per clip, better on image-to-video by a margin that survives its own error bars, and probably better on text-to-video and editing on evidence that does not yet separate it from 1.1. It carries a safety filter that removes part of an editing benchmark and will remove part of your editing workload, and it inherits 1.1's traffic on a date G​oogle has already published.

The comparison that gets quoted will be the rank movement, because twelve to six reads like a jump and forty-two Elo reads like a number. The comparison that predicts how next quarter goes is the one with the intervals on it, and that one says: one clear win, two probable ones, and a cheaper meter underneath all three.