Omni 1.2 vs. 1.1: 31 cents less per clip, and one line served twice

Our tests from 05/11/2026 on Google's official Gemini API: 9 clips at 720p, one take each, cost calculated from the tokens the API returned.

Google swears its new AI video model costs 30% less than the old one. Me, I wanted the receipt, not the press release. So I paid for both: 9 clips, $8.33 by my math. One 10-second clip: about $0.70 on Omni 1.2, $1.02 on Omni 1.1. Make 100 of them: $70 vs. $102. The discount is real. So is the freebie: a line of dialogue served twice and a word that doesn't exist.

In short

• The price, measured: a 10 s clip at 720p cost us between $0.704 and $0.706 with Omni 1.2, and between $1.016 and $1.022 with Omni 1.1. Make 100 of them: about $70 vs. $102.
• Google's receipt holds up: 5,792 tokens per second of video, exactly as on the pricing page. The extra (the model's thinking, the prompt we sent) stays under 1.5 cents per clip.
• The flubs: Omni 1.1 embellishes the café lines and flips its fox around mid-leap, while Omni 1.2 at 15 s doubles a line and invents a word. And an apple logo nobody asked for shows up on White Rose's laptop.

The 05/11 logbook

18:13 UTC. First submission. Same prompt for both versions, 720p, landscape. Omni 1.1 caps out at 10 seconds, so each scene ran three times: 1.1 at 10 s, 1.2 at 10 s, 1.2 at 15 s. Three traps:

• the fox: it runs through the snow, jumps a log, turns back toward the camera. Movement, not a word;
• White Rose, with her three reference photos: she types, stops, turns to us and says "Seven cents a second. Let's see if it's worth it.";
• the café: two women on a terrace, three lines in French at 10 s, four at 15 s, including "So let's make a hundred!".

The exact prompts, in English, are at the bottom of the page. 18:21: the ninth clip drops. Eight minutes, $8.33, under the $13 ceiling we'd set ourselves.

The receipt, line by line

Google doesn't bill by the second, it counts tokens: 5,792 per second of 720p, says its pricing page. I opened the API response for all 9 clips. Not one exception: 57,920 video tokens for 10 seconds, 86,880 for 15. A million video tokens is worth $12 on 1.2, $17.50 on 1.1. The rest of the receipt (the model's thinking, at the text rate, and the prompt we sent, White Rose's photos included) weighs less than 1.5 cents per clip.

Tokens read from the API response; cost calculated at the official prices of 05/11/2026.

Against the announced prices Kira had pinned down in her Omni 1.2 tab, the average of our 3 clips lines up: $0.705 for the roughly $0.70 announced (1.2, 10 s); $1.019 for roughly $1, i.e. ten times $0.10 a second (1.1, 10 s); $1.054 for roughly $1.04 (1.2, 15 s). The discount holds: at 10 s, the 1.2 clip costs almost 31% less than the 1.1 clip, that's 31 cents off $1.02.

Say you're putting together a hundred little 10-second ads, one per product. On 1.1, you come out at about $102. On 1.2, about $70. The $31 gap pays for 44 more clips on 1.2. And on December 3, 2026, the question goes away: Google redirects all 1.1 requests to 1.2, at 1.2's price.

What I saw (and heard)

The fox first: run, jump, look at the camera, all three beats are there in all three clips. Except that in 1.1, around 6 s, mid-leap, its head and tail melt into each other, and it lands facing the other way.

White Rose next: recognizable in all three clips, line spot-on word for word. But look at her computer.

An apple logo. Nobody asked for it. In 1.1, it stays visible for all 10 seconds; in 1.2 at 15 s, from 0 to about 4 s; and in 1.2 at 10 s too, at an angle, for the first three seconds. In our clips, both versions copy a real brand. Fine.

The café, finally. This is where it gets ugly. Each clip went through two transcribers, Gemini 3.8 Flash and Gemini 3.1 Pro Preview, each on its own, same instructions. Not Whisper, my usual second opinion: it wasn't installed on our machine that day, and the attempt failed.

• 1.2 at 10 s: all three lines, exactly. Nothing to say.
• 1.1 at 10 s: it embellishes. Around 0 s, "Did you see the price there…" (different ending depending on the transcriber), then a "Hm?" around 1 s that nobody had written, and around 5 s, "So let's make a hundred!" becomes a question.
• 1.2 at 15 s: the flub of the evening. Around 5 s, the third line starts echoing: "For a fifteen-second clip, that's one dollar. It's one dollar." And around 9 s, the last one ends on an invented name: "at Jonadi" for one transcriber, "at Genadi" for the other.

Wait, the background. The prompt only mentioned two women. In 1.1, the terrace is packed: customers behind them, and a waiter in an apron crossing the back from about 6 to 9 s. In 1.2 there are passersby too, but far off, in front of empty tables. And in 1.1, the chalkboard outside the café reads "L'ARÉSE," a word that doesn't exist.

The full clips, with sound:

▶ Play the video

My verdict, in glasses of milk

One caveat, but a big one: 3 scenes and one take per clip is not a study. Just what our 9 clips showed.

• Omni 1.2 at 10 s: 31 cents cheaper per clip, and the only one of our café clips with no dialogue errors. 🥛🥛🥛🥛
• Omni 1.2 at 15 s: about $1.05, but the doubled line and the invented word ruin the café. Re-listen before you publish. 🥛🥛🥛
• Omni 1.1: pricier, it embellishes, and it disappears on December 3. 🥛🥛

What we don't know yet

• Chance: another take might double a line at 10 s, or not at 15 s.
• Other resolutions: everything is in 720p. 360p, 1080p and 4K weren't shot.
• Google's statement: our costs are calculated from the tokens the API returns and the public prices, not read off an invoice.

Key terms

• Token: the unit Google counts to bill you; one second of 720p is worth 5,792 tokens.
• Thinking tokens: what the model "thinks" before producing the video, billed at the text rate.
• 720p: an image 1,280 by 720 pixels, the resolution of our clips.
• Transcriber: an AI that writes down what it hears in a video.
• Take: one generation; two takes of the same prompt give two different clips.

Source passages

Tests from 05/11/2026 between 18:13 and 18:21 UTC; pricing page and model card captured on 05/11/2026 at 17:15-17:16 UTC (sources of our article no. 122). Amounts in US dollars.

Show the 10 sources

Ref.	Quote (original language)	URL
S1	"We have dropped the price 30% from Gemini Omni 1.1 making it available at $0.07 per second for 720p video." ; "You can now generate and edit videos natively up to 15 seconds long, up from 10 seconds with Omni 1.1."	https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemini-omni-1-2-flash/
S2↑1↑2↑3↑4↑5↑6	(Omni 1.2 Flash) "Input price" "$1.50 (text / image / video)" ; "Output price (including thinking tokens)" "$9.00 (text) $12.00 (video)" ; "calculated at a rate of 5,792 tokens per second of 720p video. Under Standard pricing ($12.00 per 1M tokens), this equates to an effective price of approximately $0.07 per second (~$0.70 for a 10s clip, ~$1.04 for a 15s clip"	https://aistudio.google.com/docs/pricing#gemini-omni-1.2-flash
S3↑1↑2↑3↑4	(Omni 1.1 Flash) "Effective December 3, 2026, all API requests to gemini-omni-1.1-flash automatically route to gemini-omni-1.2-flash and are billed at gemini-omni-1.2-flash rates." ; "$9.00 (text) $17.50 (video)" ; "this equates to an effective price of approximately $0.10 per second."	https://aistudio.google.com/docs/pricing#gemini-omni-1.1-flash
S4↑1↑2	"Output token limit 65,536 tokens" ; (1.2) "Output video 3s-15s (360p/720p/1080p/4K, 24 FPS" ; (1.1) "Output video 3s-10s"	https://aistudio.google.com/docs/models/gemini-omni-flash
R1↑1↑2↑3↑4↑5↑6+2	Usage field of each API response (models gemini-omni-1.1-flash and gemini-omni-1.2-flash, status completed): video tokens 57,920 (10 s) and 86,880 (15 s), i.e. 5,792 per second for all 9 clips; thinking 278 to 1,450; input 117 to 187 (text only), 3,356 to 3,379 including 3,192 for images (White Rose); cost per clip and total $8.3337	— (in-house test)
R2↑1↑2↑3	Exact prompts. Fox 10 s: "In a single continuous tracking shot, no scene cuts: a red fox runs through a snowy pine forest at dawn, low camera following at its side. [0-4s] the fox sprints between the trunks, snow puffing from its paws. [4-7s] it leaps over a fallen log, its tail streaming behind. [7-10s] it lands, slows to a trot and looks back toward the camera, breath visible in the cold air. Natural forest sounds and the crunch of snow, no music, no dialogue. Realistic wildlife documentary look." White Rose 10 s: "The young woman shown in the three reference images (same face, light blue eyes, long blue hair, pointed elf ears clearly visible, pink beanie) sits at her desk in a Parisian apartment at night, warm desk lamp, a laptop open in front of her. In a single continuous shot, no scene cuts. She wears an oversized grey knit sweater. [0-3s] she types, then stops and frowns at the laptop screen. [3-7s] she turns to the camera and says in French, with a natural Parisian accent: « Sept centimes la seconde. On va voir si ça vaut le coup. » [7-10s] she raises one eyebrow and gives a small mischievous smile. Quiet room tone and keyboard sounds only, no music. Realistic, natural lighting." Café 10 s: "In a single continuous medium shot, no scene cuts: two women in their late twenties sit at a small table on a Paris café terrace in the afternoon, two coffee cups between them. One has short dark curly hair, the other a long auburn braid. [0-3s] the curly-haired woman looks at her phone and says in French: « Tu as vu le prix ? » [3-6s] the woman with the braid answers in French: « Sept centimes la seconde. » [6-10s] the curly-haired woman laughs and says in French: « Alors on en fait cent ! » They both laugh. Natural café ambience, no music. Realistic." Café 15 s, added line: "[6-10s] the curly-haired woman says in French: « Pour un clip de quinze secondes, ça fait un dollar. » [10-15s] the woman with the braid smiles and says in French: « Alors on en fait cent ! » They both laugh and clink their cups." (15 s versions of the fox and White Rose: same scenes, timings re-cut)	— (in-house test)
R3↑1↑2↑3↑4↑5	Transcriptions, two models run separately with the same instructions. Café 1.1 10 s: « T'as vu le prix là-haut ? » / « T'as vu le prix là, toi ? » ; « Hum ? » / « Hmm ? » ; « Alors on en fait cent ? » (00:05). Café 1.2 10 s: the 3 lines. Café 1.2 15 s: « Pour un clip de 15 secondes, ça fait 1 dollar. C'est 1 dollar. » (00:05) ; « Alors on en fait 100 ? Chez Jonadi ? » (gemini-3.8-flash, 00:09) / « Alors on en fait 100 chez Genadi. » (gemini-3.1-pro-preview, 0:09). White Rose, 3 clips: « 7 centimes la seconde. On va voir si ça vaut le coup. »	— (in-house test)
R4↑	Whisper attempt of 05/11/2026: "ModuleNotFoundError: No module named 'numpy'" (no Whisper transcription produced)	— (in-house test)
R5↑1↑2↑3↑4↑5	Milky's calculation R1: average of the 3 clips at 10 s, 1.2 = (0.7037 + 0.7062 + 0.7042) / 3 = $0.7047; 1.1 = (1.0179 + 1.0220 + 1.0163) / 3 = $1.0187; × 100 = $70.47 vs. $101.87, gap $31.40; (1.0187 − 0.7047) / 1.0187 = 30.8% of the average cost of a 1.1 clip; 31.40 / 0.7047 = 44.6, i.e. 44 more 1.2 clips for the same budget; 1.2 at 15 s = (1.0558 + 1.0566 + 1.0510) / 3 = $1.0545; surplus over the video portion of $1.0426: $0.0084 to $0.0140	— (in-house calculation)
R6↑1↑2↑3↑4↑5↑6+2	Frame-by-frame review of the 9 clips (contact sheets at 1 frame per second, full-resolution crops): apple logo on the lid in 1.1 10 s (0 to 10 s), 1.2 15 s (0 to about 4 s) and 1.2 10 s (0 to about 3 s, at an angle; seen at 0, 1.5, 2.75 and 3 s, out of frame at 3.25 s); café 1.1: customers at neighboring tables, waiter in an apron in the back from about 6 to 9 s, chalkboard outside the café "L'ARÉSE" (made-up word, visible from 0 to 9 s); café 1.2: passersby in the distance, empty tables; fox: run, jump, look at the camera in all 3 clips; fox 1.1: between 5.8 and 6.3 s (review at 12 frames/s), mid-leap, head and tail melt into each other, the fox runs left before and lands facing right	— (in-house test)