Gemini Omni 1.2 Flash: Which AI Video or Image Model Fits Each Business Job?

A typical marketing request this week looks like this: one product photo, one set of brand guidelines, and a list of five outputs. A hero image for the product page. Forty ad variants in every size the channels want. A 15-second social video with a voice-over. A short explainer built from the spec sheet. A looping GIF for the launch email.

Two days ago, every one of those outputs would have needed a slightly different answer. That is still true, but the answers changed. OpenAI released GPT Image 2.6 on 4 November, and today Google made Gemini Omni 1.2 Flash generally available, with 15-second clips and a 30% price cut.

Our view: there is no best AI video generator or best AI image generator for business. There is a best model for each deliverable, and the cheapest good choice for a hero image is almost never the cheapest good choice for a 15-second ad. Teams that pick one tool for everything either overpay on volume work or underdeliver on the pieces that carry the brand.

In short:

• Gemini Omni 1.2 Flash is now the obvious default for short product and social video: 15 seconds, editable by conversation, $0.07 a second at 720p and $0.104 at 1080p.

• GPT Image 2.6 Sunburst is the pick for images where layout and text must be exactly right. Its predecessor tops both of Artificial Analysis' image boards.

• Nano Banana 2.1 is the volume engine for image variants, at $0.0168 per 1K image in batch.

• Wan 3.0 is the one to watch for turning documents into video, but access is still rolling out.

• OpenAI has no video API. Sora 2 was removed on 24 September; the new animated GIF output is the closest it gets.

What Gemini Omni 1.2 Flash changes

The headline numbers are length and price. Clips now run up to 15 seconds, up from 10 in version 1.1, set in one-second steps from 3 to 15. Google says it has "dropped the price 30%." Its pricing page bills video by output tokens, which works out to about $0.07 a second at 720p, roughly $1.04 for a full 15-second clip.

The workflow detail matters more than the price. Google's intended pattern is to draft at 360p or 720p and upscale only the take you keep. At 360p, a 15-second draft costs about 35 cents, so ten rough versions cost less than three finished 1080p clips.

Three other changes make it more useful for business video than for novelty clips:

• References. You can supply character, product or style images, motion clips of up to three seconds, and audio cues. Google's Cloud announcement says the model holds products, characters and brand details more consistently across shots. Audio references are gated behind allowlisting for now.

• Conversational editing. You can swap a product or background, change the lighting or switch the aspect ratio across the full 15 seconds without regenerating the clip. For a team cutting one ad into vertical, square and widescreen versions, that removes most of the rework.

• Speech. Google describes clearer lip-sync and distinct voices for multiple speakers.

The early customers are telling. Photoroom uses it for its Video Ad Maker and Product Animator and says it chose the model for the balance of "quality, speed, and cost." QuickFrame AI says 15 seconds at 1080p lets users create "a complete CTV ad in a single generation." Both are short, commercial, repeatable jobs, which is where this model fits.

Google is also clear about the limits: fast sports action, rapid script changes and complex physics are "still in development." Every output carries an invisible SynthID watermark, which matters for any team that has to show where its content came from.

What GPT Image 2.6 adds

OpenAI's update is quieter but broader. The API changelog lists two models, GPT-Image-2.6 Sunburst and Flare, "with adjustable complexity, SVG generation, and animated GIF generation," plus markup-guided editing, where you draw on an image to show what should change.

OpenAI's image generation guide positions Sunburst "for detailed visual work" and Flare "for fast iteration." The new complexity setting controls how much the model plans before rendering. Low suits focused edits such as changing a background; high helps with "dense text, complex layouts, and relationships between objects." OpenAI also warns that higher complexity "can increase latency and introduce unwanted changes," so it is worth testing per task. The setting has no separate fee, and token prices are unchanged from 2.5.

In ChatGPT, users can now make animated GIFs and save references to people, pets and characters for reuse. OpenAI says Flare can return an image "in as little as six seconds" and that generation is 50% faster overall, both company claims we have not tested.

On quality, the 2.6 models are too new for independent boards. Their predecessors set the bar: GPT Image 2.5 Sunburst ranks first on Artificial Analysis' text-to-image leaderboard and first on its image editing board. It is also the most expensive model near the top, at about $210 per 1,000 images at maximum quality, roughly six times Nano Banana 2.1.

One gap is worth stating plainly. OpenAI removed the Sora 2 models and its Videos API on 24 September 2026 and listed no replacement. If your workflow needs video through an API, OpenAI is not currently an option.

The best model for each business visual

Working through the brief from the opening, at list API prices on 5 November:

A few notes on the less obvious choices.

Images: pay for Sunburst only where the image carries the brand. A product page hero, a campaign key visual or anything with legible copy in the frame justifies the premium. Forty resized variants of an approved visual do not. Nano Banana 2.1 is described by Google as having "accurate text rendering, and search-grounded generation across 1K, 2K, and 4K," and its batch rate halves the standard price. Generating those 40 variants costs about 67 cents.

Video: the leaderboard and the price list disagree. On Artificial Analysis' text-to-video board, the top-ranked model, Wan 3.0, costs $12 a minute, while Seedance 2.5, ranked fifth, costs $34.12. HeyGen Video 1 sits within 20 points of the leader at $2.40 a minute, though at 768p. Gemini Omni 1.1 Flash ranks tenth; version 1.2 is not on the board yet, so its quality claims are Google's until it is.

Product motion: decide what "good enough" is worth. Seedance 2.5 leads the image-to-video board, which is the closest test of turning a product photo into a clip. It costs more than five times as much per minute as Gemini Omni 1.2 at 1080p. For a hero campaign that difference may be worth paying. For a catalogue of 500 product spins it almost certainly is not.

Documents to video: promising, not yet routine. Wan 3.0 accepts web pages, PDFs and PowerPoint files and "produces video that reflects the source's structure," at up to 30 seconds. That fits internal training clips and product explainers. Its weights are closed, its access is still staged, and its document-to-video claims have not been independently tested.

Presenter-led training video is a separate category. If the requirement is a person explaining a policy on camera, a general video model is the wrong tool. HeyGen, for example, pairs its video model with a separate avatar model built from a recording of a real person. Treat that as its own buying decision, with consent and likeness rights settled first.

The generation call is the cheap part

Add up the brief and the list cost is about $8.44, plus the GIF. Nobody's budget breaks on that number. What costs money is everything around it: finding the right product photo and the approved copy, checking each output against brand rules, getting legal sign-off on claims, resizing for each channel, and keeping a record of which model made which asset with which prompt.

This is the same pattern we see with language models. In our posts on choosing a Claude model and on open-weight models like Mistral Large 4, the conclusion was that the model is a decision per step, not per company. Visual work follows the same rule. The hero image, the variants and the video are different steps with different quality bars, and each one deserves the model that clears its bar at the lowest cost.

The steps around the generation call are where workflows break. On Beam's agent platform, teams define the goals, allowed actions, tool access and escalation paths for each agent, with human handoffs where judgement is needed and an audit trail of what happened. In a visual workflow, that means the brief is pulled from the right system, the output is checked before it goes anywhere, a person approves anything customer-facing, and the record exists when someone asks later where an asset came from. The model behind the generation step will keep changing, as this week showed. The process around it should not have to.

Common questions about Gemini Omni 1.2 Flash and AI visual models

What is Gemini Omni 1.2 Flash?

Gemini Omni 1.2 Flash is Google's video generation and editing model, made generally available on 5 November 2026. It creates clips of 3 to 15 seconds with native audio, supports image, motion and audio references, and lets you edit a finished clip by conversation. It is available through the Gemini API, Google Cloud, Flow, Google Vids and Google Ads.

How much does Gemini Omni 1.2 Flash cost?

On Google's pricing page, video output works out to about $0.07 per second at 720p, $0.023 at 360p, $0.104 at 1080p and $0.209 at 4K, with the higher resolutions produced by upsampling. A 15-second 1080p clip costs about $1.56. Google says that is 30% less than version 1.1.

What is the best AI video generator for business ads?

For short product and social ads, Gemini Omni 1.2 Flash offers the best mix of length, editing and price this week. For the highest image-to-video quality, Seedance 2.5 leads Artificial Analysis' board at about five times the cost. For low-cost volume at 768p, HeyGen Video 1 and MiniMax H3 rank near the top of the text-to-video board.

What is the best AI image generator for realistic product photos?

GPT Image 2.5 Sunburst ranks first on Artificial Analysis' text-to-image and image editing boards, and GPT Image 2.6 Sunburst is its successor. For high volumes of variants, Nano Banana 2.1 ranks fourth on both boards at about a sixth of the price.

Does OpenAI still offer a video generation API?

No. OpenAI removed the Videos API and the Sora 2 models on 24 September 2026 and listed no replacement. GPT Image 2.6 can produce animated GIFs, which covers simple loops but not video with sound.