Testing AI Video Variants on Meta After the Flexible Format Went Away

Quick answer:Meta's help page says "Starting in March 2026, the flexible format will no longer be available in Ad setup." Nothing replaced it one for one. The ad that held up to ten videos and let delivery pick between them is gone, and the feature with the similar name, flexible media, does a different job: it takes one video and pushes it into more placements. What you have left for a batch of AI video variants is three ordinary tools. A multi-ad ad set ranks hooks cheaply but forfeits Meta's creative fatigue diagnostics, which the help page says are "only available for ad sets with one creative except those with Advantage+ catalogue ads (previously dynamic ads), dynamic creative or Meta Advantage+ app campaigns", and which are "not available with the sales objective before an ad set is active". One-creative ad sets keep those diagnostics but fragment budget. And the A/B test tool is the comparison Meta's A/B pages endorse, at a recommended minimum of seven days, a hard maximum of thirty, with an audience used nowhere else. What follows is how to lay twenty AI renders across those three.

What went away, and what it never was

The flexible format page is short and worth reading in its own words. "When you select Flexible as your ad format, you can select up to 10 images and videos in a single ad campaign, and the ad delivery system will automatically determine what media or media combination, such as single image, video or carousel, to show to people." It was restricted to the traffic, engagement, sales and app promotion objectives, and its third listed benefit was that grouping media with different text combinations would "help prevent ad fatigue".

It is the second multi-asset container Meta has pulled back from the sales and app promotion objectives in under two years. The dynamic creative page still carries this note: "As of June 2024, you may no longer be able to use dynamic creative when creating ad sets on Ads Manager when you select sales or app promotion as your objective." That page recommended the flexible format as the successor. One sentence on that page applied just as well to the flexible format, and applies now to whatever you build in its place. "Because results are shown as the aggregate performance across all variations, using dynamic creative as a substitute for split testing is not recommended."

The flexible format was a delivery container, not a testing tool. Losing it changes how you package variants; it does not take away an experiment, because the container never gave you one.

Flexible media is not the flexible format

Meta's flexible media page carries an explicit disambiguation: "Flexible media is different from the flexible ad format. Flexible media determines if a media asset could be used across additional placements to help improve performance, while the flexible ad format looks at multiple media assets uploaded for an ad and determines which ad format is most likely to help improve performance."

Four other sentences on that page decide how you should treat it with AI video.

• "Flexible media may be turned on by default but you can turn it off at any time." Assume it is on until you have looked.
• "For all other objectives and conversion locations, flexible media is gradually moving to Advantage+ creative." The two exceptions are the awareness objective and any objective with calls as the conversion location, which keep the toggle on the Crop screen. On a sales campaign, if the toggle is not on the Crop screen, it is on the Enhancements screen as Flex media, in the same list as the AI-labelled enhancements. The Advantage+ creative page confirms the pattern for that whole list: "Some enhancements may be turned on by default, but you can turn them off at any time."
• "We aren't currently able to detect if objects or text have been cropped out of videos like we're able to for images, but you can use ad previews to better understand how your videos may appear before you publish your ad." A 9:16 talking-head render with a burned-in hook line at the top of frame can lose that line in a Feed crop, and Meta will not catch it.
• "This product or feature may not be available to you yet." Two accounts can see different toggles this month.

So flex media is not where your variant test lives. It is a per-asset setting that multiplies placements for one render.

The one comparison Meta endorses

Meta's A/B testing page describes the mechanism: "We show each version to a segment of your audience and ensure that nobody sees both, then determine which version performs best." It measures "on a cost per result basis or cost per conversion lift basis". And it is blunt about the alternative: "We do not recommend testing informally, such as by turning ad sets or campaigns on and off manually. This can lead to inefficient ad delivery and unreliable test results."

The best-practices page adds the constraints that matter for planning. "For the most reliable results, we recommend a minimum of 7-day tests. A/B tests can only be run for a maximum of 30 days, but tests shorter than 7 days may produce inconclusive results." You should "test only one variable", and "you shouldn't use this audience for any other campaign that you're running on Facebook, Instagram or other Meta technologies at the same time."

Read those constraints against a stack of twenty renders and the A/B tool is a two-arm, single-variable, week-long instrument. It fits a structural decision and not a ranking of twenty hooks. Does an AI actor batch beat the creator footage you already run? Does a 15-second Kling 3.0 cut beat an 8-second Veo 3.1 cut of the same script? Those are two-arm questions. The informal version of that ranking, twenty ads in an ad set with the losers switched off by hand, is the thing Meta says it does not recommend.

Where Meta's fatigue signal works, and where it goes dark

The creative fatigue recommendations page opens with its eligibility rule: "This feature is only available for ad sets with one creative except those with Advantage+ catalogue ads (previously dynamic ads), dynamic creative or Meta Advantage+ app campaigns. It is not available with the sales objective before an ad set is active."

When it is available, the two delivery statuses have published thresholds. "When cost per result is more than ads that you've ran in the past, but less than twice as much, you will see a Creative limited status. When cost per result is more than or equal to twice as much as ads that you ran in the past, you will see a Creative fatigue status." The exposure count behind them is not confined to the ad set: "We consider all recent exposures of the ad's image or video, including those from other campaigns from your Page."

Two of Meta's listed remedies are worth quoting because they contradict the reflexive move. The first: "Create a new ad with a new image or video that is materially different from the original creative." The second, attached as a note: "Keeping your original ad active instead of pausing or turning it off may maximise results." Add, do not swap.

This is the trade the flexible format used to hide. Pack many videos into one ad set and the fatigue diagnostics switch off. Run one creative per ad set and they come back, at the price of budget spread across ad sets that share an audience. Neither layout is wrong; they answer different questions.

What "materially different" means when the videos are generated

The phrase comes from Meta's help page, but the reason it matters is in Meta's own research. The 2023 paper from Analytics at Meta tracks exposures "at the level of the creative rather than the ad" and says why: "If the creative is the same (or very similar) across multiple ads, then the user will likely fatigue from seeing both ads." The paper's instruction is not to cap repetition but to spread it: it says repetition "is optimally distributed across a large number of different looking ads". The exposure counts and the fitted decay, and the refresh cadences blogs attribute to them, are in creative fatigue: what Meta's own study says.

"Different looking" is the load-bearing phrase for AI creative. Twenty scripts against one actor image, one kitchen background and one opening frame produce twenty files and, to a system counting exposures per creative, something much closer to one. The variables that make a batch read as separate creatives are the ones a buyer holds constant for consistency: the actor, the setting, the first frame, the product shot. Vary those and you have twenty creatives; vary only the sentence the actor says and you have one creative with twenty captions. What makes AI UGC look fake covers the tells that repeat across renders.

A layout that uses all three tools for what they are

1. Decide structure with an A/B test, once. Two arms, one variable, seven days or more, an audience reserved for the test. The variable is the thing you cannot learn from delivery: AI actor batch against existing creator footage, or one model's cut against another's. Both arms should contain the same number of creatives so the test is about the creative source, not the count.
2. Rank hooks inside one ad set, and treat it as ranking. One ad set, one budget, let delivery allocate, and read the result as "which of these earned spend" rather than as a controlled comparison. Do not switch losers off by hand in the first days. Our sample-size guide covers how many variants that batch needs before a winner means anything.
3. Move survivors into one-creative ad sets for scale. That is the configuration where Creative limited and Creative fatigue can appear, and where the "materially different" replacement rule has something to attach to. The testing framework we run has the same shape: ranking pass, kill rule, scale pass with fresh variants queued.
4. Check flex media and the AI enhancements on every survivor. The toggles may be on by default, previews are the only cropping check Meta offers for video, and the AI-labelled enhancements can change the ad after you approved it. The full list of 31 enhancements and the two places to switch them off are in the Advantage+ creative post. Labelling is a separate topic; Meta's AI-generated creative policy covers what gets labelled and what gets rejected.

What a batch of genuinely different variants costs

The flexible format's disappearance stings because generation had become cheap enough to fill it. Here is what twenty single-clip variants cost in credits on our own catalog today, at the priced tier closest to a 10-second hook on each model. These are our prices, not a benchmark, and they move when vendors move theirs; the cost benchmarks post carries the full per-model table.

Three models are missing on purpose. Pruna Avatar, OmniHuman 1.5 and VEED Fabric are billed per spoken second of the script rather than by a fixed tier, so a batch on those depends on how long each script runs. Seedance 2.0 runs on BytePlus rather than Replicate in our production pipeline; it is still sold in fixed credit tiers (285 credits for 10 seconds at 720p, 635 at 1080p today), but it is blocked on the free trial and its 1080p tier is priced above every row in this table, so it is left out of a hook-ranking batch.

The credit column is only half the budget. Twenty different-looking variants means twenty actor and setting choices before any render starts, and that preparation, not the credits, decides whether the fatigue system sees twenty creatives or one.

Sources

• Meta Business Help Centre, "About the flexible ad format"
• Meta Business Help Centre, "About flexible media in Meta Ads Manager"
• Meta Business Help Centre, "About dynamic creative in Meta Ads Manager"
• Meta Business Help Centre, "About Advantage+ creative"
• Meta Business Help Centre, "About A/B testing"
• Meta Business Help Centre, "Best practices for A/B testing"
• Meta Business Help Centre, "Tools to create A/B tests on Meta technologies"
• Meta Business Help Centre, "About creative fatigue recommendations in Meta Ads Manager"
• Analytics at Meta, "Creative Fatigue: How advertisers can improve performance by managing repeated exposures" (10 May 2023)

All nine were read directly on 4 September 2026. Meta's help pages change without a changelog, so if a quoted sentence no longer appears on the linked page, the page moved before this post did.

The batch is only worth structuring if the variants are worth running. Twelve hook formulas is the place to start when the twenty scripts are still blank, and how to make UGC ads with AI walks the render itself. If the twenty are going to be twenty creatives rather than twenty captions on one face, they need different actors, settings and first frames, and that is the part UGC Vids AI is built for: several models from one credit balance, from $49 a month, free for the first 3 days.