Migrating From Gemini Omni 1.1 Flash to 1.2 Flash

Migrating from Gemini Omni 1.1 Flash to Gemini Omni 1.2 Flash: renamed fields, four endpoints, 15 s clips, a ~30% lower bill and what to re-test.

Gemini Omni 1.1 to 1.2 Flash migration guide: a film editor comparing two takes

Searches for Gemini Omni 1.1 Flash have been the breakout query around Google's Omni video family in Google Trends for the last 90 days. A lot of people built on 1.1. Now Gemini Omni 1.2 Flash is on Segmind, with clips up to 15 seconds and a lower bill, and every one of those integrations has the same question: what breaks if I switch?

I moved my own test harness over this week and treated it like any other dependency upgrade. I fired the same prompts, seeds and input images at both versions, 19 clips in total, and diffed the parameters, the bills, the frames and the audio. This post is the migration guide I wanted before I started: the parameter changes, the measured price change, and the list of things worth re-testing before you flip production traffic.

Short answer: it is not a version bump in a string. One endpoint became four, two fields were renamed, the duration limit now depends on which inputs you send, and seeds do not carry over. The good news is that 1.2 tells you about most of it with clear errors, and every clip I made cost about 30% less.

Gemini Omni 1.1 Flash to 1.2 Flash: what changed at a glance

Gemini Omni 1.1Gemini Omni 1.2 FlashWhat you do
Endpointsgemini-omni-1.1 plus gemini-omni-1.1-video-editFour task endpointsRoute each call by its inputs
First / last frame fieldsfirst_frame, last_framefirst_frame_image, last_frame_imageRename
Max duration, text-to-video10 s15 sOptional: use the extra 5 s
Max duration, first frame only10 s15 sOptional
Max duration, first + last frame10 s10 sNothing
Max duration, reference images10 s10 sNothing
Reference images on editNot a fieldUp to 10New capability
5 s clip at 720p, billed$0.537$0.375About 30% cheaper
Seed 42, same promptOne takeA different takeRe-approve golden takes
Resolutions / aspect ratios / audio360p to 4K, 16:9 or 9:16, always-on stereo audioSameNothing

One endpoint becomes four

On 1.1, a single endpoint, gemini-omni-1.1, did text-to-video, image-to-video, first-to-last-frame interpolation and reference images, depending on which fields you filled in. Editing lived on its own endpoint, gemini-omni-1.1-video-edit. Gemini Omni 1.2 Flash splits the generation modes apart, so the first job of a migration is a small router:

Your 1.1 request containsSend it toRequired field
Only a promptgemini-omni-1.2-flash-text-to-videoprompt
first_frame, optionally last_framegemini-omni-1.2-flash-image-to-videofirst_frame_image
reference_imagesgemini-omni-1.2-flash-reference-to-videoreference_images (1 to 10)
A video to editgemini-omni-1.2-flash-video-editvideo

Everything else is unchanged: prompt, resolution (360p, 720p, 1080p, 4k), aspect_ratio (16:9 or 9:16) and seed keep their names and values. Every clip I pulled from both versions came back as H.264 at 24 fps with a 48 kHz stereo AAC track, and audio still cannot be switched off.

Gemini Omni 1.2 Flash parameter changes, field by field

Field1.11.2 Flash
duration3 to 10 s on every mode3 to 15 s on text-to-video and first-frame image-to-video; 3 to 10 s once last_frame_image is set; 3 to 10 s on reference-to-video; rejected on edit
first_frameImage URLRenamed first_frame_image
last_frameImage URLRenamed last_frame_image
reference_images1 to 10, not combinable with frames1 to 10, on reference-to-video and on video-edit
aspect_ratio on editNot a fieldRejected: output keeps the input's aspect ratio

The duration rule is the one to read twice. A first-frame clip can now run 15 seconds, but the moment you also pass a last frame, the ceiling drops back to 10. A 1.1 job that sends both frames at 10 seconds migrates cleanly; a new feature that adds a last frame to an existing 15-second job will start failing.

The errors you will hit, and the one you won't

I mapped the parameter space with deliberately invalid requests, which return a 400 and cost nothing. Gemini Omni 1.2 Flash is unusually helpful here. Send it a 1.1-shaped payload and it names the fix:

# first_frame on any 1.2 endpoint
{"error": "`first_frame` is not a field on this model; use `first_frame_image`"}

# an image on the text-to-video endpoint
{"error": "`first_frame_image` is not supported by gemini-omni-1.2-flash-text-to-video; use gemini-omni-1.2-flash-image-to-video instead"}

# reference images on the image-to-video endpoint
{"error": "`reference_images` is not supported by gemini-omni-1.2-flash-image-to-video; use gemini-omni-1.2-flash-reference-to-video or gemini-omni-1.2-flash-video-edit instead"}

# a last frame with duration 11
{"error": "duration must be between 3 and 10 seconds"}

# duration on the edit endpoint
{"error": "duration is not supported by gemini-omni-1.2-flash-video-edit; the output length follows the input video"}

So the renamed fields fail loudly, which is what you want. The quiet failure is everything the validator does not know about. A field name that is neither current nor a known 1.1 name passes straight through and the call runs on defaults. Earlier the same day, one stray parameter in a test request on 1.2 came back as a default 5-second clip and billed $0.38. Typos cost money; they do not error.

The same applies in reverse if you ever roll back. 1.1 did not reject first_frame_image at validation; it moved straight on to checking the other fields. I did not pay to prove what it does with the image, but a rollback that keeps the 1.2 field names should be treated as a text-to-video call until you have checked. Keep both payload shapes in code, not in a find-and-replace.

Here is the change for the most common call, an image-to-video transition:

import requests

# Before: Gemini Omni 1.1
requests.post(
    "https://api.segmind.com/v1/gemini-omni-1.1",
    headers={"x-api-key": "YOUR_API_KEY"},
    json={
        "prompt": "Time-lapse with a locked-off camera: ...",
        "first_frame": "https://your-bucket.s3.amazonaws.com/dawn.jpg",
        "last_frame": "https://your-bucket.s3.amazonaws.com/night.jpg",
        "duration": 8, "resolution": "720p", "aspect_ratio": "16:9", "seed": 42,
    },
)

# After: Gemini Omni 1.2 Flash
resp = requests.post(
    "https://api.segmind.com/v1/gemini-omni-1.2-flash-image-to-video",
    headers={"x-api-key": "YOUR_API_KEY"},
    json={
        "prompt": "Time-lapse with a locked-off camera: ...",
        "first_frame_image": "https://your-bucket.s3.amazonaws.com/dawn.jpg",
        "last_frame_image": "https://your-bucket.s3.amazonaws.com/night.jpg",
        "duration": 8, "resolution": "720p", "aspect_ratio": "16:9", "seed": 42,
    },
)
resp.raise_for_status()
open("transition.mp4", "wb").write(resp.content)

Gemini Omni 1.2 Flash pricing vs 1.1, measured on the same prompts

Every number below is what Segmind billed me for that exact call on November 5, 2026, same prompt and seed on both sides.

CallGemini Omni 1.1Gemini Omni 1.2 FlashChange
5 s, 360p$0.182$0.130-28.6%
5 s, 720p$0.537$0.375-30.2%
5 s, 1080p$0.804$0.559-30.5%
5 s, 4K$1.604$1.110-30.8%
10 s, 720p, text-to-video$1.072$0.744-30.6%
8 s, 720p, two reference images$0.862$0.598-30.7%
8 s, 720p, first + last frame$0.860$0.598-30.5%
Edit a 10 s clip at 720p$1.378$0.834-39.5%
15 s, 720p, text-to-videonot possible$1.112 to $1.115new

At 720p that works out to about $0.107 per second on 1.1 and $0.074 per second on 1.2 Flash. Editing saved the most, almost 40%. One thing worth knowing if you budget from model pages: the 1.1 page lists about $0.64 for a 5-second 720p clip, but I was billed $0.537, so your real 1.1 baseline may already be lower than the table you planned against. Compare your own invoices, not the two tables.

For a concrete number, take the 1.1 page's own example: 30 five-second 720p clips a day. At what I was billed, that is about $483 a month on 1.1 and $337 on 1.2 Flash, roughly $146 a month back for changing an endpoint and two field names.

What to re-test before you switch

Price and parameters are the easy part. These are the behaviours I would check against your own prompts before moving production traffic, with what I saw on mine.

1. Seeds do not carry over

Prompt usedA single continuous shot: a red vintage bicycle leans against a whitewashed wall on a narrow Mediterranean street, pink bougainvillea moving in the breeze; a grey cat walks past the front wheel and out of frame. Ambient sound: breeze, a distant church bell, a scooter passing. No music, no on-screen text, no logos.

Parametersduration: 5  |  resolution: 720p  |  aspect_ratio: 16:9  |  seed: 42

Gemini Omni 1.1

Gemini Omni 1.2 Flash

Same prompt, same seed 42, 720p. 1.1 billed $0.537 in 38 s; 1.2 Flash billed $0.375 in 33 s.

Both versions understood the scene, but they are different takes: frame correlation between them was 0.08 to 0.14, which is effectively unrelated. If you approved a 1.1 take by its seed and regenerate it on 1.2, you get a new shot. Store the approved 1.1 files, and re-approve seeds on 1.2 before anything downstream depends on them.

2. Your 720p draft now previews your 1080p and 4K render

Gemini Omni 1.1
720p, t = 2.5 s

Gemini Omni 1.1 output at 720p: bicycle and cat example frame

Gemini Omni 1.1
1080p, t = 2.5 s

Gemini Omni 1.1 output at 1080p: bicycle and cat example frame

Gemini Omni 1.2 Flash
720p, t = 2.5 s

Gemini Omni 1.2 Flash output at 720p: bicycle and cat example frame

Gemini Omni 1.2 Flash
1080p, t = 2.5 s

Gemini Omni 1.2 Flash output at 1080p: bicycle and cat example frame

Same prompt and seed at two resolutions. On 1.1 the bicycle turned around between 720p and 1080p; on 1.2 Flash the two frames are the same render.

I rendered the bicycle prompt at all four tiers on both versions and correlated frames at five timestamps. On Gemini Omni 1.2 Flash, 720p, 1080p and 4K scored 0.999 against each other at every timestamp: one render delivered at three sizes. On 1.1, 1080p and 4K were the same render (0.998 to 0.999), but 720p was a different take (0.68 to 0.74). When I ran the same test on 1.1 in August, its 720p matched its upper tiers, so treat this as behaviour that can move under you rather than a fixed rule.

The practical change: on 1.2 Flash, iterate at 720p and only pay 1.49x for 1080p or 2.96x for 4K when the delivery spec needs the pixels; the composition will not move. 360p is still its own take on both versions (0.57 to 0.63 on 1.2), so it is a cheap idea sketch, not a preview.

3. Timecoded prompts, and what happens when you just bump the duration

Prompt used[0-3s] Close-up of a potter's hands centring wet clay on a spinning wheel. [3-6s] The camera pulls back to show the potter, a woman in a mustard apron, at the wheel in a sunlit studio as the clay rises into a tall cylinder. [6-10s] She lifts her hands away and the finished vase keeps spinning as the camera pushes in on it. Ambient sound: the hum of the wheel, wet clay, birdsong outside. Single continuous shot, no cuts. No music, no on-screen text.

Parametersduration: 10  |  resolution: 720p  |  aspect_ratio: 16:9  |  seed: 42

Gemini Omni 1.1

Gemini Omni 1.2 Flash

The same 10-second timecoded prompt. 1.1 billed $1.072; 1.2 Flash billed $0.744.

Both kept a single unbroken take, as asked; scene detection found no cuts in either. 1.2 Flash was closer to the marks: the pull-back to the potter had happened by about 3.6 seconds, where 1.1 stayed on the hands until around 5. 1.2 also gave the clay a terracotta tone and a shapelier vase, so if your look depends on 1.1's palette, check it.

Then the lazy migration that most people will try first: keep the 10-second prompt and set duration to 15.

Old timecodes, duration 15

Rewritten timecodes, duration 15

Gemini Omni 1.2 Flash at 15 seconds. Left: the unchanged 10-second prompt. Right: beats rewritten as [0-5s] [5-10s] [10-15s]. Both billed about $1.11.

Prompt used[0-5s] Close-up of a potter's hands centring wet clay on a spinning wheel. [5-10s] The camera pulls back to show the potter, a woman in a mustard apron, at the wheel in a sunlit studio as the clay rises into a tall cylinder. [10-15s] She lifts her hands away and the finished vase keeps spinning as the camera pushes in on it. Ambient sound: the hum of the wheel, wet clay, birdsong outside. Single continuous shot, no cuts. No music, no on-screen text.

Parametersduration: 15  |  resolution: 720p  |  aspect_ratio: 16:9  |  seed: 42

It did not pad, freeze or loop. The model stretched the three beats across the full 15 seconds: the close-up ran past its 3-second mark, the pull-back took the middle, and the push-in on the finished vase arrived around 12. Motion stayed active to the last second. The rewritten prompt landed each beat near its new mark. Either works, but I would still rewrite the timecodes, because with old marks you are trusting the model to reinterpret your timing rather than following it.

4. Reference images and <IMAGE_REF_n> tags

Reference image 1
<IMAGE_REF_0>: the barista

Reference image for Gemini Omni 1.2 Flash: barista portrait

Reference image 2
<IMAGE_REF_1>: the dripper

Reference image for Gemini Omni 1.2 Flash: sage-green ceramic coffee dripper

Both stills made on Nano Banana 2 and passed in reference_images, in this order, to both versions.

Prompt used<IMAGE_REF_0> stands behind a cafe counter and slowly pours hot water from a gooseneck kettle into <IMAGE_REF_1>, steam rising, then looks up at the camera and says: "Thirty seconds of bloom. Then you pour." Ambient cafe sound, soft clatter of cups. No music, no on-screen text, no logos.

Parametersreference_images: 2  |  duration: 8  |  resolution: 720p  |  aspect_ratio: 16:9  |  seed: 42

Gemini Omni 1.1

Gemini Omni 1.2 Flash

Reference-to-video with the same two images and tags. 1.1 billed $0.862 in 46 s; 1.2 Flash billed $0.598 in 70 s.

The tag syntax carries over unchanged: both versions bound <IMAGE_REF_0> to the barista and <IMAGE_REF_1> to the dripper, and both spoke the line. I ran each audio track through a speech recogniser and both came back as "30 seconds of bloom, then you pour." The difference was product fidelity. 1.1 turned the dripper into a sage-green cup on a saucer; 1.2 Flash kept it a dripper, with a filter, the gold rim and the ring handle. If you use reference mode for products, that is the strongest reason to migrate I found. It was also the one case where 1.2 was slower, 70 seconds against 46.

5. First and last frame: the motion now happens earlier

First frame
first_frame / first_frame_image

First frame input for Gemini Omni 1.2 Flash image-to-video: studio at dawn

Last frame
last_frame / last_frame_image

Last frame input for Gemini Omni 1.2 Flash image-to-video: studio at night with lamp on

The same two Nano Banana 2 stills, sent under the 1.1 names to 1.1 and the 1.2 names to 1.2 Flash.

Prompt usedTime-lapse with a locked-off camera: daylight in the ceramics studio fades from dawn to night, the sky outside darkens and windows in the houses light up one by one, then the brass desk lamp clicks on and warms the bowls. Ambient sound: birdsong fading to crickets, a soft click of the lamp switch. No music, no on-screen text.

Parametersduration: 8  |  resolution: 720p  |  aspect_ratio: 16:9  |  seed: 42

Gemini Omni 1.1

Gemini Omni 1.2 Flash

First-to-last-frame transition. 1.1 billed $0.860; 1.2 Flash billed $0.598. Both took 45 s.

Both versions lock the ends equally well: the first frame correlates 0.98 with my still and the last frame 0.99 on each. The timing is what changed. I measured frame-to-frame motion every half second. 1.1 spreads the transition over the clip and clicks the lamp on at around 6 to 7 seconds. 1.2 Flash finishes the dusk and the lamp by about 5 seconds, then holds an almost still frame for the last 2. If your 1.1 transitions are cut to music or timed to land at the end, re-time them: ask for a shorter duration, or put the timing in the prompt.

6. The edit endpoint, and the new reference_images field

Prompt usedChange the potter's apron to deep indigo blue. Keep everything else the same.

Parametersvideo: the 10 s Gemini Omni 1.1 potter clip above  |  resolution: 720p  |  seed: 42

gemini-omni-1.1-video-edit

gemini-omni-1.2-flash-video-edit

The same source clip and instruction on both edit endpoints. 1.1 billed $1.378; 1.2 Flash billed $0.834.

Both made the apron indigo and left everything else alone; frames correlate 0.96 to 0.99 with the source on both. I could not pick a winner on quality, so on edits the migration is simply the 39.5% saving, plus 1.2's longer inputs: the 1.1 edit page caps input at 10 seconds, and in my launch-day tests the 1.2 edit endpoint took a 15-second clip and returned 15 seconds.

Edit on 1.2 Flash also accepts reference_images, which 1.1's edit endpoint does not have. I tried it on the same clip:

Prompt usedReplace the finished vase on the wheel with the sage-green ceramic dripper with the gold rim from <IMAGE_REF_0>. Keep the potter, the camera motion and everything else the same.

Parametersvideo: the same 10 s clip  |  reference_images: 1 (the dripper)  |  resolution: 720p  |  seed: 42

Gemini Omni 1.2 Flash video edit with a reference image: the vase replaced with the reference dripper. Billed $0.835.

It works, with caveats. The dripper on the wheel matches the reference by the end, gold rim and handle included, and the potter and camera move are untouched. But the clay turned sage-green from the first frame, and around the 5 to 7 second mark the model briefly stacked two drippers before settling on one. I would treat reference edits as a new feature to prototype with, not a drop-in step for a production pipeline yet.

7. What did not change

  • Output format: H.264, 24 fps, 48 kHz stereo AAC on every clip from both versions.
  • Duration accuracy: both return about one frame over the request (5.03 s, 8.04 s and 10.03 s on both; 15.04 s on 1.2), so any trim logic still holds.
  • Latency: no consistent speed-up. 5 s at 720p took 38 s on 1.1 and 33 s on 1.2; 4K took 167 s and 147 s; the 10 s clip took 48 s and 49 s.
  • The resolution and aspect-ratio enums, the seed field, and always-on audio.

A migration checklist

  1. Route each 1.1 call to one of the four 1.2 Flash endpoints by its inputs.
  2. Rename first_frame and last_frame to first_frame_image and last_frame_image.
  3. Keep frame-pair jobs at 10 seconds or less; only first-frame and text-only jobs get 15.
  4. Remove duration and aspect_ratio from edit calls.
  5. Validate field names in your own code: unknown fields bill instead of failing.
  6. Re-approve any seed you rely on, and keep the 1.1 renders you already shipped.
  7. Re-time first-to-last-frame transitions; 1.2 front-loads the motion.
  8. Move drafts from 360p to 720p, since 720p now matches your final render.
  9. Rewrite timecoded prompts if you raise the duration.

FAQ

Is Gemini Omni 1.2 Flash cheaper than 1.1?

Yes. On identical prompts I was billed 28.6% to 30.8% less at every resolution for text-to-video, about 30.6% less for reference and frame-to-frame clips, and 39.5% less for a 10-second edit.

Does my Gemini Omni 1.1 payload work on 1.2 Flash?

Not unchanged. Text-only prompts move over as-is to the text-to-video endpoint. Frame jobs need two renamed fields and a different endpoint, and 1.2 returns a 400 naming the fix if you forget.

How long can Gemini Omni 1.2 Flash videos be?

Up to 15 seconds for text-to-video and first-frame image-to-video. With a last frame, or in reference mode, the limit is still 10 seconds. Edits follow the length of the input video.

Will the same seed give me the same video on 1.2 Flash?

No. The same prompt and seed produced unrelated takes on 1.1 and 1.2 Flash. Within 1.2, the same seed and inputs are reproducible, so re-approve seeds once after migrating.

Do I need to change my reference image prompts?

No. The <IMAGE_REF_0> style tags work the same way on the reference-to-video endpoint, with 1 to 10 images. In my test 1.2 Flash also kept the product closer to its reference.

Is Gemini Omni 1.2 Flash faster than 1.1?

Not consistently. Most of my calls finished within a few seconds of 1.1, 4K was about 20 seconds quicker, and reference mode was slower at 70 seconds against 46.

Should you migrate?

If you are on 1.1 for text-to-video or edits, yes: it is a router, two renamed fields and a re-approval pass, and the bill drops by about a third. Reference-mode users get better product fidelity on top. The one place to slow down is first-to-last-frame transitions, where the new pacing will change how your cuts feel.

Try Gemini Omni 1.2 Flash text-to-video, image-to-video, reference-to-video and video edit on Segmind, and keep Gemini Omni 1.1 running side by side while you re-test. Available via API with no setup required.