How 5 Popular AI Video Generators Handled Our Tests

AI video generators are suddenly everywhere; every major platform has added a Video tab, homepages promise cinema from a prompt, and affiliate roundups keep recommending the same names while somehow finding a category for nearly all of them to win. We had seen enough of those lists, so this time we went straight to the models and made the videos ourselves.

We chose five current text-to-video AI models from Arena’s leaderboard: Seedance 2.5, Gemini Omni, FLUX 3, MiniMax H3 and HappyHorse. Each received three ten-second scenes built around weaknesses that commonly exposed older models: a father interrupting his teenage son tested dialogue, eyelines and whether a strategy game could continue inside a generated monitor; a hotel fight tested fast movement, physics and whether three people, one red gift and an elevator could remain in the same scene; while the Mango Cola advert tested attractive faces, liquid, product text, colour, native audio and commercial pacing.

We did not rerun the clips until each model finally understood us, because that would have turned the comparison into a prompting contest. The mistakes therefore stayed visible: one son answered his father while looking in the wrong direction, several fight scenes lost track of the attackers or moved the elevator, and the Mango Cola branding often changed between frames. At the same time, some of the gameplay looked surprisingly convincing and several adverts were close to usable, which gave us a fair picture of what these models can do now and where they still fall apart.

Quick answer

Seedance 2.5 wins, which isn't much of a surprise for the newest and most expensive model in this group. It was the best all-rounder across lighting, colour, people, movement and native sound; the Mango Cola result was the cleanest commercial, and its hotel fight was the only one that put the hero on the correct side of the elevator doors. It still lost one attacker during that fight, wrote “MANGOM COLA” in an opening frame, and took more than 13 minutes to produce ten seconds, so paying the most did not remove the usual AI mistakes.

MiniMax H3 came second and probably made the best father-and-son video, with a convincing middle-aged dad, active gameplay and small details such as realistic mouse movement. It also made a good advert for $1.40, although its fixed 2K output gave it a resolution advantage over most of the group and its fight was complete nonsense.

FLUX 3 finished third as the average all-rounder. Its game looked good, Mango Cola stayed unusually readable and nothing apart from the fight went seriously wrong, but it never produced the best result in any category. Gemini Omni came fourth because $1 bought us the fastest generation and a lively advert, while HappyHorse finished fifth despite making the second-closest attempt at the fight; its father-and-son scene and drink advert were weaker than the rest.

These prices were recorded during our test sessions on 1 September 2026. FLUX ran through Black Forest Labs’ own interface, while the other four models ran through WaveSpeed, and the amounts below show what these exact settings cost us rather than what every future generation will cost.

How we tested the five AI video generators

We ran the comparison on 1 September 2026 and started with Arena’s text-to-video leaderboard, where the 25 August snapshot placed Gemini Omni 1.1 Flash first, FLUX 3 Video third, Dreamina Seedance 2.5 fifth, MiniMax H3 sixth, and HappyHorse 1.0 eighth. HappyHorse 1.1 was already available when we began, so we used the newer version instead of following the table blindly.

Arena gave us the shortlist, not the final ranking. Its table contained several versions from the same model families, while newer entries naturally had fewer votes than models that had been available for longer; we therefore chose five current options that could all accept comparable ten-second text prompts. FLUX ran through Black Forest Labs, while Gemini Omni, HappyHorse, MiniMax and Seedance ran through WaveSpeed.

We had already compared ten AI image generators and four image-to-video models, but those tests either judged a single frame or gave the video model an image to begin with. This time, every model started from text alone, without a first frame, character reference, product photograph, audio track, or ending image to keep it on the road.

We did not choose three unrelated scenes because they sounded fun. Each one put pressure on a different group of problems:

1. The father and son conversation tested a door opening, two recognisable characters, natural acting, correct eyelines, spoken dialogue, and a generated strategy game continuing inside a generated computer monitor. It was the least complicated scene physically, but the model still had to create one video inside another and understand which way the son should look when his father sat beside him.
2. The hotel fight tested fast physics, three people interacting at once, body and object continuity, sound effects, a fixed room, and a required ending. The red gift needed to stay somewhere sensible, both attackers needed to remain separate people, and the elevator could not move to another wall whenever the camera turned.
3. The Mango Cola advert tested beautiful people, eyes and skin, liquid physics, fizz, condensation, an exact product name, a black-and-white-to-colour transition, commercial pacing, and a suitable native voice. The people and drink needed to look desirable, but a beautiful first frame was useless if the product changed its name by the end.

Every video was ten seconds long and 16:9, with native audio enabled where the endpoint supported it. We kept the first completed generation and did not reroll the scenes after seeing the mistakes, because a model that needs four private attempts before the people face the correct direction has not passed a first-result comparison.

MiniMax H3 is the one settings complication. Its endpoint produced fixed 2K video, while Seedance, Omni and HappyHorse ran at 720p and FLUX used HD, so MiniMax deserves the result it made but cannot receive an unqualified image-quality victory over lower-resolution outputs.

The original prompt histories were no longer available when we assembled this article. The prompt toggles below reproduce the instructions from our testing record, including the scenes, dialogue, required objects, audio and timing, but we are not presenting the punctuation as a character-for-character export from the generation interfaces.

What AI video generation cost and how long it took

One generation from each model cost $8.91 at our recorded settings, putting all three rounds at approximately $26.73 before any separate experiments. The prices were reasonable compared with the all-in-one AI video platforms we tested previously, where a $20 InVideo subscription used its entire allowance without giving us a finished video, although direct model testing becomes expensive once the rerolls begin.

The time column records one completed run per model rather than a laboratory average, and generation queues, discounts and hosting prices can all change. Omni finished in 57.77 seconds, which made it easy to experiment with, while Seedance needed around 13 minutes and 25 seconds for ten seconds of output; the quality was better, but waiting that long becomes annoying very quickly when one disappearing attacker forces another $3.24 attempt.

Test one: a father interrupts his son’s computer game

This looked like the easy test, and compared with the fight it was, but there was more going on than a father walking into a bedroom. The door had to open naturally, the same father and son needed to survive the full scene, both spoken lines had to belong to the correct person and the strategy game inside the monitor had to keep looking like gameplay rather than a colourful desktop wallpaper; after the father sat down, the son also needed to remember which side of the room he was on.

MiniMax H3 result

MiniMax H3 made our favourite version. Its father looked like a believable middle-aged dad, the door movement was clean and the wide monitor showed an impressive game that continued to feel active after he entered, right down to the son’s mouse movement. This was probably the strongest single result outside Seedance’s advert, although MiniMax had the benefit of 2K output and we are not going to pretend that extra resolution had nothing to do with how good it looked.

Seedance 2.5 result

Seedance had richer lighting and colour, the dialogue worked and its fictional game was just as convincing, but the monitor was absolutely enormous, almost like the family had installed a small cinema across the boy’s desk. It was funny rather than fatal, and the father must be doing very well if his teenage son has that setup, yet MiniMax still made the more natural room.

FLUX 3 Video result

FLUX did a respectable job without producing anything that popped in the same way. The father, boy and gameplay stayed believable, and the invented game looked good enough that we genuinely wanted to play it; higher resolution may have pushed the scene closer to MiniMax, but the tested output remained a decent third-place version rather than a winner.

Gemini Omni 1.1 Flash result

Gemini Omni was close until the conversation stopped making physical sense. The father entered from the left, said “Still awake?” and then sat beside the boy on that same side, yet the son turned to his right and shouted “We’re losing the planet!” toward an empty part of the room. The game, bedroom and people still looked good, which somehow made the mistake more obvious because everything else had convinced us that the scene understood itself.

HappyHorse 1.1 result

HappyHorse lost the round. The boy played a believable teenage son, but the father opened the door with his eyes apparently closed and carried the familiar expression of an NPC waiting for us to select a dialogue option; the strategy game also looked noticeably worse than the versions made by MiniMax, Seedance, FLUX and Omni.

MiniMax wins the first test, with its 2K advantage stated clearly; Seedance had the best lighting and atmosphere, FLUX was fine without doing anything special, Omni spoiled a good attempt by looking the wrong way and HappyHorse never caught up.

Test two: three men, one gift and an elevator

The second prompt was deliberately nasty because action exposes problems that disappear when one attractive person slowly turns toward the camera. Our hero had to walk through a fixed hotel lobby with a red present, notice two attackers arriving from different directions, put the gift somewhere sensible, fight both men and enter the elevator before the ten seconds ended; meanwhile, the attackers had to remain separate people, the furniture needed to stay put and the sound had to match whatever their bodies were doing.

Seedance 2.5 result

Nobody handled all of it. Seedance came closest because the men ran and fought with convincing speed, the impacts had some weight, the sound made sense and the red gift remained visible on the table, while the hero was also the only one who entered the elevator correctly. It still lost the plot around the attackers: one man dropped without a clear hit and the other disappeared from the back of the scene, presumably because Seedance had finished using him.

HappyHorse 1.1 result

HappyHorse made the second most realistic fight, especially in the running, impacts and action-film sound, but its people acted like NPCs who had not received the same mission. One armed man ran away, the other pointed his gun without shooting and the hero stayed outside the elevator while the doors closed in front of him, even though the framing tried to convince us he had stepped inside.

MiniMax H3 result

MiniMax was less successful and much funnier. The hero caught the first man’s gun hand, moved with the urgency of somebody tidying a table and then gave him a light tap on the head with the weapon, which somehow knocked him out; while that was happening, the second attacker tried to hit our hero and appeared to punch his own friend instead. The main character kept nearly the same expression through everything, then watched the elevator close while he was still outside. John Wick has plenty to learn from these guys.

FLUX 3 Video result

FLUX made the fight look less like an ambush and more like a confused exchange in a hotel lobby. One attacker seemed to walk over and hand the hero his gun, after which the hero placed or threw him onto the sofa for a little rest; the second man approached, saw what was happening and walked away with the completely understandable energy of “Shit, I’m out of here.” The click during the gun handover sounded more like somebody’s wrist breaking than a weapon being removed, but at least the elevator remained in roughly the correct place.

Gemini Omni 1.1 Flash result

Gemini Omni failed hardest. The gift moved around, one attacker entered another man’s body like Casper or Danny Phantom and the elevator jumped from one side of the lobby to the other while the walls changed around it, turning the hotel into some Doctor Strange version of New York architecture. Individual frames were fine, yet the full sequence had so little spatial memory that replaying it only made it funnier.

This is why the fight mattered. Every model could create a polished lobby, a man in a suit, running, punches and weapon sounds, but none could keep three bodies, one red box and a fixed room logically connected once everything began moving at the same time; Seedance failed least badly, HappyHorse was the next closest physically, and the remaining attempts worked better as comedy than action.

Test three: make Mango Cola look drinkable

The Mango Cola prompt was our business test: an invented product with an exact two-word name, an orange can, cola and mango, beautiful people, a black-and-white-to-colour transformation and a short native voiceover. A usable advert needed the people and liquid to look desirable while the product, label and basic story survived the same ten seconds.

Seedance 2.5 result

Seedance made the cleanest commercial: the people, colour, sound, condensation and can opening looked convincing, there was only one drink in the scene and the whole thing looked closest to something a real brand might use; it moved a little too quickly, but the people’s eyes were fine, the liquid looked drinkable and the native voice suited the advert. The opening label appears to say “MANGOM COLA,” with an extra M attached to Mango, so even the winner still needs its exact branding added later in an editor.

FLUX 3 Video result

FLUX surprised us because Mango Cola stayed readable, the fizz and pouring worked, the drink looked cold and the woman taking the man’s can followed a logical little story. The voice was excellent too. Its problem was the couple, especially the woman’s eyes and facial contours, which still looked obviously generated despite the HD setting; nothing was catastrophically wrong, but the people kept it behind Seedance.

MiniMax H3 result

MiniMax produced another good result for $1.40. The orange can, cola, ice and final product frame looked realistic, while the main label remained much cleaner than Omni’s; there was some invented microtext around the rim, the voice had a slightly NPC quality and other models made more of the transition from grey into colour, but the actual drink and can survived the full advert.

Gemini Omni 1.1 Flash result

Omni may have understood the language of advertising better than it understood the label. A mango dropped into the glass, an orange ring moved through the liquid and the colour transition gave the clip the pacing of a proper commercial; its people also looked more realistic than FLUX’s, apart from some wandering eye and iris movement. The product text was complete trash, however, with letters rotating, overlapping and changing position while the can moved.

HappyHorse 1.1 result

HappyHorse decided that our drink was called CONGA. Large vertical letters carried that name from the black-and-white opening into the final colour frame, while smaller versions of Mango Cola appeared around the top and bottom of the can; it failed the instruction, but at least it stayed loyal to the brand it had invented for itself. The people were stiff, the mango was difficult to understand and the fizz when the can opened was the only part of the advert that really stood out.

Seedance wins the advert test because it gave us the best combination of people, product, sound and commercial finish. FLUX and MiniMax were usable starting points, Omni made the most interesting animation around the worst text, and HappyHorse accidentally launched CONGA.

Best AI video generators in our test: the final ranking

1. Seedance 2.5

Seedance wins because it gave us the best collection of first results rather than one lucky clip. Its advert looked the closest to finished, the father-and-son scene had the strongest lighting and its fight was the only one that ended with the hero inside the elevator; the people, sound and camera movement were also more convincing across the three prompts, which is why Seedance remains our TGK favourite.

It was also the most expensive and slowest model here. One discounted ten-second clip cost $3.24 and took 805.03 seconds, while the regular price shown was $3.60; Seedance then added an extra letter to Mango Cola and lost one of the attackers during the fight, so the best result in this group still needs rerolls or a more controlled production process when the scene becomes complicated.

2. MiniMax H3

MiniMax was the strongest alternative. It won the father-and-son prompt with the most believable dad, gameplay and small physical details, then produced a Mango Cola advert that was close to usable; at $1.40 it cost less than half as much as Seedance in our session, although it took more than six minutes to finish.

The fight was hopeless and extremely funny, and MiniMax also had the advantage of fixed 2K output, which makes a perfectly equal sharpness comparison impossible. Even with that caveat, it gave us the best balance between price and quality, and it would be the first model we tried if Seedance was too expensive for the job.

3. FLUX 3 Video

FLUX was fine at almost everything and best at nothing, which makes it the average all-rounder of this test rather than a hidden winner. The fictional game looked playable, the father and son held together, the advert made the drink appealing and Mango Cola was among the cleanest labels in the group; we also liked Black Forest Labs’ uncluttered interface, although that had no effect on the video ranking.

Its fight failed, the faces in the advert still carried an obvious generated look and the HD output did not have Seedance’s finish or MiniMax’s 2K detail. FLUX charged us $1.87 and landed exactly where its results belonged: in the middle, capable of making a decent clip but without winning a reason to choose it first.

4. Gemini Omni 1.1 Flash

Omni gave us the cheapest and fastest generations: $1 and 57.77 seconds for the recorded run. Its Mango Cola advert had lively animation, a suitable voice and realistic-looking people, while the gaming scene would have competed with the leaders if the son had turned toward his father instead of shouting to the empty side of the room.

The hotel fight was the weakest action result and the product lettering broke too badly for a real advert, which is why speed and value could not move it above FLUX. We would still use Omni for cheap drafts, visual ideas and fast prompt experiments before paying Seedance prices for a final attempt.

5. HappyHorse 1.1

HappyHorse’s fifth place hides one useful strength: its hotel scene contained the second most convincing physical action after Seedance, while the running, impacts and sound were better than the FLUX, Omni or MiniMax versions. It also kept its accidental CONGA branding surprisingly consistent, which is some kind of achievement even if it was not the achievement we asked for.

That could not rescue the full set. HappyHorse made the weakest game, gave us the most NPC-like father and produced a stiff advert with the wrong brand; it might deserve another action-focused attempt with slower and clearer choreography, but every other model gave us a stronger reason to choose it for a general text-to-video job.

What we would use after this test

We would choose Seedance 2.5 when the decent clean look matters most and the budget can survive a few failed generations. MiniMax H3 is more reasonable when $1.40 and fixed 2K output suit the job, especially for a slower scene involving one or two people, while Gemini Omni is the easy choice for cheap drafts and quick visual experiments; FLUX sits between them as an average all-rounder with reasonably good product text, but this test did not give us a clear reason to pick it ahead of Seedance or MiniMax for final footage.

For any complicated action, we would change the production method before changing the model. The hotel fight should be divided into several short shots, with reference images used to establish the hero, attackers, gift and lobby; important branding should also be added afterward in an editor, because even Seedance could not keep two invented words completely correct. That involves more work than one long prompt, but it removes many of the opportunities to move the elevator, merge two people or delete an attacker halfway through the scene.

If you want one service to write the script, choose the voice and assemble the final edit, see our separate test of five free AI video makers. Those platforms solve a different problem and come with a very different kind of pricing nonsense.

Final verdict

Seedance 2.5 remains our favourite text-to-video AI generator and wins this comparison, while MiniMax H3 came close enough to be the better value choice for many jobs. FLUX 3 was average across the three prompts, Gemini Omni traded consistency for excellent speed and price, and HappyHorse showed some ability with physical action without matching the others in the gaming scene or advert.

The fight was the most useful result because nobody passed it. These models can already generate a convincing computer game, attractive people, a cold drink and individual punches that look real, but three moving characters still created disappearing attackers, wandering elevators and broken cause and effect; for that kind of scene, shorter generated shots and a normal editor remain much safer than asking one model to understand the whole sequence in ten seconds.