I Made the Same AI Video Ad 9 Ways

Halfway through opening a soda can, the monkey pulled a second pull tab out of thin air.

Google's Veo 3.1 had been asked, in those exact words, for "the tab stays attached." It delivered a tab that stayed attached — plus a spare, in case of emergencies. That shot cost 34 cents.

Over three days in October I made the same soda ad nine ways: no AI at all, free models on my local desktop, two paid models, and a mix.

The short version of what I learned has nothing to do with which model to buy. A small business can make a usable AI video ad for pocket change. The model matters less than the source material you feed it and how carefully you check what comes out. And the failures in the blooper reel had me laughing out loud.

*Our tool sent no mode, so OpenArt's default applied. Today that default is Lite, and OpenArt keeps no per-clip record of the mode, so Lite is our best inference.

Why make a monkey open a can?

The soda is called BANANA BOOM and doesn't exist in the real world, despite the intrigue a banana-flavored soda brings to my taste buds. The brief sounds trivial, but it packs four things into five seconds that turned out to be exactly where things broke: hands, a tiny mechanism, liquid going where it should, and a crack that has to land on the exact frame.

The AI doing the work was Claude Code, Anthropic's coding agent: it wrote the scripts, ran every model, cut the edits and mixed the sound, and separate Claude agents reviewed each result. I wrote the brief, watched every cut and decided what shipped. When Claude got something wrong, I'll say so by name.

Level 0: no AI at all

Monkey #1 is drawn in code with Remotion, and every sound was synthesized from scratch. Twelve minutes to build, 38 seconds to render, $0. The "ahhh" was robotic, because math is a poor substitute for a satisfied primate.

For a logo sting or a mascot that must look identical every time, code still wins: the brand name never misspells itself.

Level 1: a 3D cartoon, free, on my desktop

The first real AI version ran entirely on my own desktop, with two free models doing the picture work, both licensed for commercial use:

Then the monkey drank the soda, and the soda went down his chest.

So the next take's prompt said, in so many words, "no spilling and no dripping."

It still spilled.

Claude then asked for four short clips to bridge the gaps, each one told "completely dry: no liquid, no dripping, no spilling." All four dripped. A product shot with no monkey in it at all poured soda over a can that was still closed. The only dry take was the one where he didn't drink.

Two drinks, four bridges, six leaks. I've never seen an instruction ignored with such commitment.

We shipped it anyway, by cutting around the leaks. A lot of AI video editing is exactly that: rescuing the usable frames.

Level 2: photoreal, still free

Then the brief went realistic, with the monkey finding the can first, which is why everything from here runs 18 to 27 seconds.

#3, the first free photoreal cut, cast three different monkeys in one ad: one with a black cap and a peach face, one with coarse brown fur, one with a slate-blue face and a white ruff. The hand opening the can melted into a fingerless mitten.

On an empty jungle shot, Claude's prompt still mentioned the can, so Wan grew a giant, misspelled one in the forest, like a monument to the prompt. And in another take with no monkey, a human hand reached in and took the can. Somebody else filming on set wanted a soda too...

The fix had nothing to do with the video model. Claude rewrote the starting pictures to describe the same monkey, hands included, in every shot, and three species became one.

Then for #7, Qwen-Image-Edit built every picture with the monkey in it from one reference image, so he stayed the same monkey from shot to shot. Wan drafted each shot at 480p and SeedVR2 upscaled it to 1080p.

Same monkey in every shot, a finger that lifts the tab like a lever, a wide "AHHH." His hands look like rubber gloves, but model fees were still $0.

Lesson one: the starting pictures decide more than the video model does.

Level 3: what does $6 of paid AI video buy you?

Both paid models came through OpenArt, a site that resells credits for many models; its Plus plan is $34 a month for 12,000 credits. Claude gave both models the identical starting pictures and prompts, though not the same settings: Veo ran at 1080p in four-second shots, Kling in its standard 720p mode in five-second shots.

Veo 3.1 (from Google) cost $0.34 a shot, $2.38 for seven. It grew the spare tab and acted the "ahhh" without voicing it: mouth open, a lick of the lips, silence.

One caveat: Veo most likely ran in Lite mode (see the note under the first table), the lowest of OpenArt's three Veo 3.1 settings, so it probably didn't run at its best.

Kling 3.0 (from Kuaishou) cost about $0.50 a shot, $3.47 for seven. A finger hooks the ring, the tab rises, the spray comes out of the top, and the crack lands right on it.

Then the monkey's palm turns magenta, as the can's color bleeds into his hand. Apparently the paint was still wet.

Two flaws they shared were ours.

The monkey changes coats between shots (jet black in one shot, a small brown juvenile in another, a cream mane in a third), and the can-opening shot has a grey hand with flat, human-looking fingernails. Both were already in the starting pictures, and both models faithfully animated what they were handed. Lesson one again, from the paid side.

Can the AI do the sound too?

Both paid models make sound along with the picture. Neither voiced the "ahhh," the one sound the brief asked for by name.

Kling's own audio was near-silent on three of seven shots, and in the shot that introduces the can, where Claude's prompt asked for stream water, a "glint" chime and "no music," Kling added a sustained musical tone. Interesting addition...

Veo's own splash I called "laughably bad," and part of the joke was on me: Claude's audio processing had squashed it.

The sound that won was the one we built: every effect placed on a measured frame. Our own "ahhh" came from an open voice model, Chatterbox, and a reviewer measured its pitch: an adult man's voice coming out of a small monkey.

So, as a tenth step, we swapped in real recordings from the free Sonniss sound library for the jungle, the birds and the "ahhh." More believable, still too low for a small monkey, and now I could hear the last fakes: two synthesized metal "tinks" when the monkey's hand grabs the can. A palm on an aluminum can doesn't ring like a bell.

Level 4: the winner was a hybrid

The version I'd stand behind, despite the remaining flaws, is #9: the free #7, with one shot swapped for Kling's can-opening and its own crack and fizz. The paid part is one five-second shot, about 50 cents. My note when I watched it: "the best one so far."

It also has the flaw I'm most embarrassed by.

Look at the image at the top of this page: the monkey opening the can has a cream torso. A second later, drinking, he has black arms. Same monkey, two wardrobes.

Lesson two: pay only for the hard shot. Free models were good enough for the scenery, the can on its rock and the drink. The paid model was clearly better at a hand working a tiny mechanism, so that's the shot worth paying for.

What it really cost, in money and time

*As above.

So all fourteen paid shots came to (drumroll please, though you've already seen it twice)

$5.85.

Measured from the session logs, the Claude work behind the nine versions came to about 1,200 model calls: the main session plus 24 review and research agents. At Anthropic's published API rates, that would have cost about $120.

I didn't pay that. It all ran on the $100-a-month Claude Max plan I already had, so the Claude side added $0 in cash.

That $120 also isn't the price of one ad. It covers all nine versions and a lot of trial and error, and along the way Claude wrote the scripts and we worked out a process the next video can reuse. So the next one should need less Claude work, though I haven't measured how much less.

An AI agent wrote the scripts and drove every step. By hand it would take longer, and I can't tell you how much because I am not a professional video editor.

Lesson three: check it like it's trying to fool you

From the 3D cartoon on, every file went to a reviewer before I signed off, and they caught things BEFORE they were shipped: Veo's second tab, which we cut down to just three frames in the final cut; audio running 52 milliseconds late; a "recorded" jungle sound that turned out to be synthesized; and Claude telling a reviewer a gap in the sound was covered when it wasn't.

We've written about this with text: making got cheap and checking didn't. With video it's worse, because a defect can hide in three frames. Step through the frames at full resolution before anyone else sees them.

Where AI video stands in October 2026

The free video model I used, Wan 2.2, released in July 2025, is fifteen months old, and Alibaba's Wan team hasn't put a newer general-purpose video model up for download since; its newer ones you can only rent. Every free-route gain here came from what we fed it and the tools around it, not from a newer model.

RigidBench, a University of Cambridge preprint from August 2026, tested eight models, including Veo 3.1 (its full and Fast tiers; ours most likely ran in Lite) and Kling 3.0 (in its pro mode, not our standard one). No model led on all ten measurements, and the usual "does it look right" score ranked them almost backwards from how accurately they moved objects. Looking right and moving right are different skills. My extra pull tab is a small, silly example.

What I didn't test: OpenAI's Sora, Runway, and the AI video features in Canva and CapCut. The scope was what a small business can run on its own desktop or rent one shot at a time, so I have nothing to add about those tools and how they'd do. Sounds like a test for another day.

If this were your ad, what would I use?

No desktop and no coding agent? Use a browser reseller like OpenArt and pay per shot.

#8 was all paid ($3.47), though its starting pictures came from my desktop; on the site you'd make them with its image models or use your own photos. Commercial use of what OpenArt makes needs its Plus plan or above.

Whichever route: make one picture of your character you like, build every other starting picture from it, keep each prompt to what's actually in the frame, and check the result frame by frame. If nobody on your team has time for that, the video fees aren't the real price.

If you'd rather not spend three days learning which shots to pay for, start a no-pressure conversation. We'll tell you which route fits what you're selling, including when the honest answer is code, or a camera.

Cheers, from the boring side of the business,

P.S. Claude has since made a free fixed cut. The tinks are gone (I picked silence over real hand sounds), the "ahhh" is pitched up to suit a small monkey (I picked it blind from ten takes), and a patch covers the end of the label smear. The patch left a faint seam of its own. Of course it did. Next time: before and after, side by side.