Google's changelog dated November 5, 2026 labels Gemini Omni 1.2 Flash generally available. Its pricing page lists $12.00 per 1M video output tokens, about $0.07 per second at 720p by Google's footnote. Clips run 3 to 15 seconds. Google's API and Cloud pages differ on extension length (30 versus 40 seconds), Provisioned Throughput and default interaction retention (55 days versus 7). We tested nothing.
A token is the unit Google counts and bills in; for video output it is converted to a number of tokens per second of video, so the price per second depends on the resolution. Upsampling means the model first makes a clip and then enlarges it to 1080p or 4K. Provisioned Throughput is a way of reserving capacity ahead of time. GA means generally available; on the Gemini API access is paid-tier only, and some features are region-limited or gated.
SynthID and C2PA are two ways of marking a file as AI-made. SynthID is Google's invisible watermark; C2PA Content Credentials are a metadata standard that can travel with the file if the manifest is kept. The Gemini API and the Google Cloud platform are two routes to the model with separate documentation, and some details differ between their pages.
Basis: public Google pages read on November 6, 2026 - the Gemini API release notes, the Omni guide, the pricing, rate-limits and model pages, the Interactions API and Files API pages, the Google Cloud model page and four Google Cloud guides (Interactions, extend videos, upscale videos, Provisioned Throughput for Omni), the Google DeepMind model card published November 5 and two Google blog posts. We tested nothing and generated no video, and each number and quote belongs to Google, which published it. Our arithmetic and our recommendations are marked as ours. The release is a day old, so treat this as status as of November 6, 2026; the pages can change.
What did Google release as GA?
Google's release notes carry an entry dated November 5, 2026, titled "Gemini Omni 1.2 Flash generally available (GA)". It names the model code gemini-omni-1.2-flash and lists "up to 15-second video generation and conversational editing", "30% lower video output pricing ($12.00 per 1M tokens), and doubled paid-tier rate limits". A second entry says that from December 3, 2026 requests to gemini-omni-1.1-flash route automatically to 1.2 at the 1.2 rates. The Omni guide describes the same four-week transition window.
A GA label appears on five pages we read: the changelog; the pricing page (the model is "now generally available to developers on the paid tier of the Gemini API"); the Cloud model page ("Launch stage: GA", "Release date: November 5, 2026", "Retirement date: November 5, 2027 or later"); the Cloud blog (title: "now generally available"; body: "generally available on Google Cloud"); and the Google blog ("generally available across the Gemini API and Gemini for business"). The Omni guide carries no GA label, and the Gemini API model page lists the model as Stable, latest update November 2026. The Cloud page footnote adds that listed retirement dates refer to support in Gemini Enterprise Agent Platform and that the Gemini API "is subject to its own terms of service".
Google's blog says the model is rolling out across the Gemini API, Google Cloud, Google Vids, Flow, Google Ads and Gemini for business, and that video in the Gemini app is "rolling out to Google AI Ultra subscribers". The Cloud blog says Gemini for business is "coming soon". Those app surfaces are not what the changelog calls GA, and this article is about the developer API and the Cloud model page.
What does a clip cost?
Google's pricing page lists, for the Paid tier, $1.50 per 1M input tokens (text, image, video) and, for output, $9.00 per 1M for text and $12.00 per 1M for video. The Free tier row reads "Not available". The same table has a row "Used to improve our products" that reads "No" for the Paid tier; it links to Google's terms, which we did not open.
A footnote converts video output to seconds: "Billing is based on total output token consumption, calculated at a rate of 5,792 tokens per second of 720p video." Google then prints the per-second prices below. The 720p result is about $0.07 per second, which Google gives as about $0.70 for a 10-second clip and about $1.04 for a 15-second clip. We repeated the arithmetic (5,792 tokens x $12.00 per 1M tokens); the result is consistent with the token rate in Google's own footnote, and the figures are vendor-listed and approximate.
| Resolution | Output tokens per second (Google) | Price per second (Google's footnote) | Our check |
|---|---|---|---|
| 360p | 1,931 | $0.023 | 1,931 x $12 per 1M = $0.0232 |
| 720p (default) | 5,792 | about $0.07 | 5,792 x $12 per 1M = $0.0695 |
| 1080p, "with upsampling" | 8,688 | $0.104 | 8,688 x $12 per 1M = $0.1043 |
| 4K, "with upsampling" | 17,376 | $0.209 | 17,376 x $12 per 1M = $0.2085 |
| 1.1 Flash video output at $17.50 per 1M, 720p | 5,792 (Google's 1.1 footnote) | about $0.10 | 5,792 x $17.50 per 1M = $0.1014 |
Two claims in the same pages are Google's. The Google blog says "We have dropped the price 30% from Gemini Omni 1.1"; by our arithmetic $17.50 to $12.00 is about 31% lower. The Cloud blog says "$0.07 per second of 720p video, 30% less than Gemini Omni 1.1 Flash, with similar savings at other resolutions". We did not find on the pages we read how an extension or an edit turn is billed (the added seconds or the whole resulting clip), though the guide says "Each turn in the conversation produces a new video". For standalone upscaling, the Cloud upscale guide gives token rates by resolution pair (2,896 output video tokens per second for 360p to 720p and 720p to 1080p, 5,792 for 360p to 1080p, 11,584 for 720p to 4K) and bills them "at the same output video token price as all other Omni 1.2 Flash tasks", but prints no dollar figure; we did not open Cloud pricing. The Cloud model page links to pricing but prints no price itself. Treat the per-second figures as an estimate of video output charges only: input tokens, text output, extra generations, edits, extensions and upscaling can change a bill.
What are the limits on length, resolution and references?
From the Omni guide and the model page, in Google's words where quoted:
- Length: "any whole second from 3 to 15 seconds", 10 seconds when you do not set one; a duration over 15 seconds "returns an invalid argument error". Output is 24 FPS per the model page.
- Resolution and shape: 360p, 720p (default), 1080p and 4K, with the last two marked "(upscaled)"; aspect ratio 16:9 or 9:16.
- First and last frame: supported for 3 to 10 seconds; two frames plus a longer duration returns an error.
- Extension: the Gemini API guide says the model "uses the final 10 seconds of the input video to generate a 3 to 10 second continuation", up to 30 seconds in total, appended to the end only. Google Cloud's extend-videos guide describes a longer total: the extension is concatenated onto up to the last 30 seconds of the input video, "for a total output duration of up to 40 seconds".
- References: up to 6 reference images per request ("up to 3 images per character for up to 2 characters"); reference videos up to 3 seconds each, up to 3 for clips of 3 to 10 seconds and 1 for clips of 11 to 15 seconds, for new generation only, with audio in a reference video ignored.
- Editing uploaded video: input of 15 seconds or less, uploaded through the Files API; the Files API page says files "are stored for 48 hours". Not available in the EEA, Switzerland or the UK (see regions below).
- Large output: the guide advises
delivery="uri"for videos over 4 MB. - Not supported (guide): voice editing, audio inputs or audio references, system instructions, YouTube videos as a source, and Provisioned Throughput.
- Editing in later turns: the guide says
store=falsemeans "the generated video won't be editable in subsequent turns".
Request limits are on Google's rate-limits page: Free "Not supported"; Tier 1 8 requests per minute and 40 per day; Tier 2 16 and 200; Tier 3 40 and 2,000. Google adds that "Specified rate limits are not guaranteed and actual capacity may vary". Google's page lists Tier 2 as requiring $100 paid and 3 days since the first successful payment, and Tier 3 as $1,000 paid and 30 days since the first successful payment. The changelog says "doubled paid-tier rate limits"; neither page gives the earlier figures, so "doubled" has no baseline for us. As a simple example (our arithmetic), 40 plain 15-second generations a day on Tier 1 would be 600 seconds of new video. That is not a ceiling on video length, because the pages do not say whether edits and extension turns count as separate requests.
Where do the API and Cloud pages differ?
Google publishes two sets of pages: the Gemini API documentation (AI Studio) and the Cloud model page, with a Cloud blog beside it. We did not find which page is current on the points below, and the pages may describe two products with different scopes; we do not call either page wrong. Statements that appear only in the Cloud blog are not in the documentation, so we mark them as blog statements.
| Topic | Gemini API pages | Cloud page and blogs |
|---|---|---|
| Provisioned Throughput | Guide: "Provisioned throughput is not supported." | Cloud page: "Provisioned Throughput" Supported, with a separate Cloud page for Omni (quota windows below). Cloud blog: "offering Provisioned Throughput". Google blog: "Provisioned Throughput is available". |
| Extension length | Gemini API guide: extension up to 30 seconds in total | Cloud extend-videos guide: extension concatenated onto up to the last 30 seconds of the input, for up to 40 seconds in total; its example: extending a 60-second input by 10 seconds "produces a 40-second output video". |
| Interaction retention | Interactions overview: 55 days on the Paid tier, 1 day on the Free tier, by default | Cloud Interactions page: by default interactions are stored for 7 days (store=true); store=false opts into stateless execution and supports Zero Data Retention workflows. |
| Image count | Guide: up to 6 reference images per request | Cloud page: "Maximum images per prompt: 10". The pages do not say whether these count the same thing. |
| Output token limit | Model page: 65,536 | Cloud page: 260,640. |
| Sampling controls | Guide: system instructions, temperature and top_p are "not supported" | Cloud page: system instructions "Not supported", and also lists defaults for temperature (0.0-2.0, default 1.0) and topP (0.0-1.0, default 0.95). |
| Audio inputs | Guide: "Uploading audio inputs or audio references is unsupported in the current version of the API." | Cloud blog: "Audio references are gated behind allowlisting." Google blog lists "audio cues" among the guidance inputs. Cloud page: Audio modality "Not supported". |
| Provenance mark | Guide: SynthID watermarking on all generated videos | Cloud page: "Content Credentials (C2PA)" Supported, with no mention of SynthID. Both blogs mention SynthID; the Cloud blog does not mention C2PA. |
| Upscaling | Guide: 1080p and 4K are upscaled outputs of the resolution setting; no standalone upscaling task listed | Cloud page: "standalone video upscaling" and "Upscale videos" Supported. Cloud upscale guide: 360p input to 720p or 1080p, 720p input to 1080p or 4K; 360p straight to 4K and 1080p to 4K "not supported"; the guide says it is optimized for videos generated by Gemini Omni models and that quality "might vary" for others. It lists token rates by resolution pair (2,896 to 11,584 per second) at the same output video token price, with no dollar figure. |
| APIs and regions | Guide: all samples use the Interactions API; EEA, Switzerland and UK limits stated | Cloud page: Interactions API and Count Tokens Supported, GenerateContent, Live and Chat Completions "Not supported", Batch inference "Not supported"; availability "Global: global" only; no EEA, Switzerland or UK wording on the page. |
One more Cloud-only detail: Google Cloud's Provisioned Throughput page for Omni lists quota enforcement windows by purchase size - 2,000 seconds for 1-9 GSUs (the capacity units the page uses), 200 for 10-19, 100 for 20-39, 50 for 40-66 and 30 for 67 or more - and says they are "subject to change". For gemini-omni-1.2-flash it says a single request too large for one window (served as pay-as-you-go) arises only for 4K videos of 10 seconds or longer. The pages we read do not tie these windows to the Gemini API tier limits above.
What does Google document about provenance and regions?
The Omni guide states: "All generated videos include SynthID watermarking, which is invisible to viewers but can be detected programmatically for provenance verification." Both blogs say outputs carry an "imperceptible SynthID watermark". The Cloud model page does not mention SynthID but lists Content Credentials (C2PA) as supported. We make no claim about how reliably any mark can be detected; the pages we read give no detection figures, and we ran no detector. The DeepMind model card, published November 5, also says "SynthID watermarks enable users to verify AI-generated content", and that although the model is capable of changing people's speech when editing videos, "for now, we are restricting this capability".
On regions, the guide states that editing or extending uploaded videos is "not currently available for users in the European Economic Area (EEA), Switzerland, and the United Kingdom", and that editing or extending videos the model generated is supported there. It also says uploading and editing images containing minors is not supported in those regions, that uploading and editing images with "certain recognizable people" is not supported, and that safety filters on prompts and output "vary by region". The Cloud page lists one location, "Global: global", for availability, Provisioned Throughput and Standard pay-as-you-go. Its security-controls block lists data residency, CMEK, VPC-SC and AXT, and sends the reader to a Security controls page we did not open.
On storage, the Gemini API Interactions overview says interactions are stored by default (store=true), retained for 55 days on the Paid tier and 1 day on the Free tier, with a Paid-tier window of 7, 14, 28 or 55 days and a delete method. Google Cloud's Interactions page gives a different default: "By default, the Interactions API stores interactions on the server (store=true) for 7 days", with store=false for stateless execution and Zero Data Retention workflows. The Cloud page lists Omni 1.2 Flash among models the Interactions API supports, but neither page's retention sentence names generated Omni video or reference images, other than the store=false editing note above. For uploads, the Files API page says files "are stored for 48 hours" and "automatically deleted after 48 hours". Both pages describe stored interactions in general (the API overview: "all Interaction objects"; the Cloud page: "interactions", for 7 days); neither says whether that covers generated video or reference images, which may be covered elsewhere. We do not know which default applies.
What do the pages not say?
Silence on a page is not a denial; these items may sit on pages we did not open (Google's terms, the Cloud pricing and security pages, the Gemini Omni Flash model card that the DeepMind card refers to).
| Topic | What the pages state | What we did not find on the pages we read |
|---|---|---|
| Status | Changelog, Nov 5, 2026: GA; pricing page: "now generally available" on the paid tier; Cloud page: Launch stage GA, release date Nov 5, 2026; both blogs | A GA label in the Omni guide; a date on the pricing, rate-limits or guide pages |
| Price | $12.00 per 1M video output tokens; token rates per second for four resolutions; 30% drop (Google's figure) | How extension and edit turns are billed; a dollar figure for standalone upscaling (the Cloud guide gives token rates and the same output video token price, and we did not open Cloud pricing); batch, caching or priority prices for Omni |
| Rate limits | Tier 1 to 3 request limits; "doubled paid-tier rate limits"; not guaranteed | The earlier limits, so no baseline for "doubled" |
| Quality | Google's blog describes consistency and motion qualities and names remaining weak spots (fast sports action, rapid script changes, complex physics) | A benchmark, an independent evaluation or an error rate; the DeepMind model card (published November 5) covers a frontier-safety note and mitigations, refers to the Gemini Omni Flash card for known limitations and evaluation (which we did not open), and carries no benchmarks |
| Provenance | SynthID on all outputs (guide, blogs); C2PA listed (Cloud page) | Detection reliability; a reconciliation of SynthID and C2PA |
| Retention | Interactions API default storage, which the API and Cloud pages state differently (see the differences table); uploads via the Files API: 48 hours (Files API page); store=false note | Whether stored interactions cover generated video or reference images; both pages describe stored interactions in general |
| Data use | Pricing page: "Used to improve our products" is "No" for the Paid tier | The terms behind that row, which we did not open |
| Service terms | Cloud blog carries a customer quote mentioning SLAs | An SLA for the model on any Google page we read |
| Current page | Both sets of pages carry the model | Which of the API and Cloud pages is current on the differences above |
Is this computer vision?
Not in the sense of inspection or detection. Gemini Omni 1.2 Flash is positioned for video generation and editing. The Cloud model page also lists a maximum input video length for video understanding of 10 minutes, while marking "Agentic video understanding" and "Computer use" as "Not supported". No source we read presents it as a replacement for detection or inspection pipelines or says it produces labelled data, and we do not suggest using it as training data for such models. This article is about costing and governance of generated video (price per second, limits, regions, provenance marks, retention gaps), aimed at operations and product leaders. Waveo's computer-vision work, which reads real footage and images, is a different tool; our teardown of Gemini Computer Use covers a different Google product that acts on screens.
One adjacent point: a generated clip is not a record of a real asset. Our note on Genie 4 and digital twins makes the same distinction for a generated world and a twin linked to an asset's data. For how Google's cloud and API offers sit beside other vendors, see our comparisons of enterprise AI agent platforms and OpenAI on Azure, AWS and Google Cloud, and the AI product teardowns hub.
Do this before a video budget
Three steps that need no developer access.
- Write down how many seconds of video you expect to make per month and at which resolution, then multiply by Google's listed per-second rate (vendor-listed, approximate: about $0.07 at 720p, $0.104 at 1080p, $0.209 at 4K) to get a baseline output-cost estimate (video output only, not a cap). Edit and extension turns each make a new video, and input tokens, text output and extra generations are not in that figure.
- Decide whether you need 1080p or 4K, clips over 10 seconds, or editing or extension of uploaded video, and read the upsampled rates and the EEA, Switzerland and UK limit before you commit.
- Ask your legal contact whether reference images and video of people or products may be sent to Google, and get retention and data-use answers in writing: the pages we read describe stored interactions in general and do not say whether that covers generated video or reference images (uploads via the Files API: 48 hours).
What should you ask before budgeting?
These are our recommendations, not Google's, for a team that wants video from this model in ads, product pages or internal training videos.
- Which of the two sets of pages governs our contract: the Gemini API or Google Cloud? The differences table above shows where the answer matters. Get it in writing.
- How many seconds of video a month, at which resolution, and how many edit or extension turns per clip? Each turn produces a new video, and the pages we read do not say how those are billed.
- Which tier will we be on, and are its request limits enough? Tier 1 is 40 requests a day on the page we read.
- Do we need clips over 10 seconds, 1080p or 4K, extension past 30 seconds (the API guide and Cloud guide differ), or editing or extension of uploaded video? Per the guide, editing or extending uploaded videos is not currently available in the EEA, Switzerland and the UK (editing or extending videos the model generated is supported).
- Who owns the check that a generated clip is labelled as AI-made wherever we publish it, and how will we verify the SynthID or C2PA marks on our side?
- May reference images and video of people or products be sent to Google, and for how long are generated video and reference images kept? The pages we read describe stored interactions in general and do not say whether that covers either; the Files API page gives 48 hours for uploaded files.
If you want help scoping a pilot or an integration, see our LLM integration service.
What is still unknown?
The pricing, rate-limits and guide pages print no date, so we cannot say when they last changed; they reflect what was public on November 6, 2026. We did not run the model and have no data on output quality, speed or failure rates. We read the DeepMind model card only for the points above; Google's terms, the Cloud pricing and security pages and the Gemini Omni Flash model card it refers to were not opened. The blog posts include customer quotes and video samples chosen by Google, which we did not evaluate. We also do not know which retention default, if either, applies to generated video. Prices, limits and pages may change, as the day-old release shows.
Sources
- Google AI Studio documentation: Gemini API release notes, entry of November 5, 2026, read November 6, 2026
- Google AI Studio documentation: Generate and edit videos with Gemini Omni Flash, read November 6, 2026
- Google AI Studio documentation: Gemini Developer API pricing, Gemini Omni 1.2 Flash and 1.1 Flash sections, read November 6, 2026
- Google AI Studio documentation: Rate limits, Gemini Omni Flash rate limits and usage tiers, read November 6, 2026
- Google AI Studio documentation: Gemini Omni Flash model page, read November 6, 2026
- Google AI Studio documentation: Interactions API overview, data storage and retention section, read November 6, 2026
- Google AI Studio documentation: Files API, read November 6, 2026
- Google Cloud documentation: Gemini Omni 1.2 Flash model page, last updated 2026-11-06 UTC, read November 6, 2026
- Google Cloud documentation: Interactions API, state management and retention, read November 6, 2026
- Google Cloud documentation: Extend videos, read November 6, 2026
- Google Cloud documentation: Upscale videos, read November 6, 2026
- Google Cloud documentation: Provisioned Throughput for Gemini Omni models, quota enforcement windows, read November 6, 2026
- Google DeepMind: Gemini Omni 1.2 Flash model card, published November 5, 2026, read November 6, 2026
- Google blog: Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips, November 5, 2026, read November 6, 2026
- Google Cloud blog: Gemini Omni 1.2 Flash now generally available, November 5, 2026, read November 6, 2026