AI video in campaigns: count cost, model changes and approval before you scale

AI video in campaigns sounds like a creative shortcut at first. Write a prompt, get a clip, publish it. That is the easy part. The hard part starts when someone wants ten variations, the same visual style next month, a clean model switch later, and a clear answer to who approved the final version.

This week gave us a useful reminder from both xAI and Google. xAI made Grok Imagine Video 1.5 generally available in the Imagine API. Google added firm shutdown dates for older Veo and Imagen model IDs in the Gemini API. That is not just creator news. It is a practical warning for any organization that wants to use AI video more than once.

What changed: faster Grok video, but not without planning

xAI describes Grok Imagine Video 1.5 as an image-to-video model with better motion, physics, audio, and sync. It is now GA in the Imagine API as grok-imagine-video-1.5, and xAI is rolling out Video 1.5 Fast in Grok on web and mobile. In xAI's example, Fast generates a 6-second 720p clip in about 25 seconds, down from more than 40 seconds before.

That sounds like a straight productivity win. Read it operationally instead: when a team can make more clips faster, it also needs better rules for which images may be used, which versions get saved, which costs are acceptable and who stops a clip before it reaches a campaign.

Source: xAI: Grok Imagine Video 1.5

The xAI model card is worth reading before anyone plans a campaign around the wrong input. It lists image and video modalities and explicitly says the model does not support text-to-video right now. Current API pricing is 0.08 dollars per second for 480p and 0.14 dollars per second for 720p, plus 0.01 dollars per image input. That is not expensive per clip. It becomes expensive when the process has no guardrails.

Source: xAI model card: grok-imagine-video-1.5 and xAI pricing

Count the campaign, not just the clip

A 10-second 720p clip with one image input costs about 1.41 dollars using the current API prices. Twenty variations cost about 28.20 dollars. That is still affordable for many campaigns, but only if someone has decided what a variation is for.

Otherwise, AI video becomes the same old problem as image generation: everyone tries a few things, nobody names the files, nobody remembers the prompt, and the version that actually worked sits inside someone's personal account.

Set one simple rule before you scale: every generated clip needs a purpose. It might test a hook, audience, format, language or channel. If nobody can say what the variation tests, do not generate it.

Google's reminder: media workflows break when model IDs die

On June 15, Google added shutdown dates for several media-generation models in the Gemini API. Veo IDs such as veo-2.0-generate-001, veo-3.0-generate-001 and veo-3.0-fast-generate-001 shut down on June 30, 2026. Imagen 4 IDs such as imagen-4.0-generate-001, imagen-4.0-ultra-generate-001 and imagen-4.0-fast-generate-001 shut down on August 17, 2026.

For a marketing team, this does not only mean switching models later. It means templates, scheduled jobs, no-code flows, prompt libraries and internal instructions can break if nobody owns the model calendar.

Source: Gemini API release notes

That matters even more for organizations that do not have a developer group watching API changelogs every week. If AI video starts showing up in training, ads, social posts or internal communication, model versions should be treated like any other dependency: named, documented and replaceable.

A practical AI video rollout check

Do not start with the prettiest film. Start with a small process you can operate.

• Source image and rights: Which image may the model animate? Is it owned, licensed or customer-owned? Can it be used in an ad?
• Model and version: Write down the exact model ID, endpoint, resolution, duration, and date. Also save why you chose it.
• Cost per approved clip: Do not count only the clips that worked. Count every attempt needed to get one publishable result.
• Brand boundary: Decide what the model must never change: logo, product shape, medical claims, school information, pricing, or customer promises.
• Approver: One person must be able to say yes or no to the final version. Not a chat channel. Not the whole team.
• Fallback: What happens if the endpoint has errors, the model changes or an old ID shuts down? Pause the campaign, switch provider or use a static image?
• Archive: Store the prompt, input image, output, cost, model ID, approver and publishing location in one place.

The last point sounds boring. It is often the difference between a fun experiment and a repeatable workflow.

Endpoint errors are not a side note

On June 17, xAI had short incidents with increased errors on the Image Generation endpoint in both us-east-1 and eu-west-1. They were resolved and lasted a little over twenty minutes. That is not a reason to avoid the tool. It is a reminder that media production through APIs needs the same calm incident thinking as other automation.

If a campaign has to go out at 09:00, you should know in advance whether you wait, switch model, lower resolution or publish an already approved backup.

Source: xAI status us-east-1 and xAI status eu-west-1

When it is worth automating

AI video works best when the process already has a frame: recurring campaign formats, training clips, internal instructions, local service ads or versions for different channels. If every clip is a new art project, automation may add more friction than help.

Hammer can help through Tool Forge when you want to turn this into a real workflow: prompt library, folder structure, model calendar, cost limits, approval steps and logging. Mindset Forge fits better if you are still deciding which clips AI should be allowed to create at all.

Start with one campaign type. Set a maximum budget for test variations. Pick one person who owns the final version. Then AI video becomes a production tool, not another folder of half-finished clips.