Introduction
In May we compared the four AI video models that mattered most at the time: Veo 3.1, Sora 2, Grok and Seedance. Veo came out as the most consistent, Sora had the highest ceiling, and both looked like safe bets for a while.
A few months is a long time in AI video. Since then a new wave of models has arrived: Google's Gemini Omni 1.1 Flash, Alibaba's HappyHorse 1.1, and two models that currently top the public leaderboards, Alibaba's Wan 3.0 and MiniMax H3. Some of them claim to beat Sora and Veo outright, and several are cheaper.
So we ran the question again, the only way that counts: same prompts, every model, one take each, and we watched every clip. This article walks through the results from several angles: speech in difficult conditions, people and food in a vertical social format, animating a still photo, and a simple talking-head line that turned out to be surprisingly revealing.
The Models
The new generation
- Google Gemini Omni 1.1 Flash. Google's newest video model, built to generate picture and synchronised sound together. Fast, up to 4K, 3–10 second clips.
- Alibaba HappyHorse 1.1. A cinematic text-to-video and image-to-video model with smooth camera work. Up to 1080p and 15 seconds, and the most expensive model in this test.
- Alibaba Wan 3.0 and MiniMax H3. At the time of testing, ranked first and fourth on Artificial Analysis's text-to-video-with-audio leaderboard, which is based on blind human votes.
The returning models
- OpenAI Sora 2. The best model in our May bar scene, and still the reference for natural speech.
- Google Veo 3.1 Lite. Our most consistent performer in May, and the best value.
- ByteDance Seedance 1.5 Pro, plus its cheaper Pro Fast variant. The budget option from May, and much cheaper than everything else here, with optional audio.
Methodology
Each test used the exact same prompt for every model, with no rewording, retries or cherry-picking. All clips were generated at 720p (MiniMax H3 at its native 768p) and kept short, 4–5 seconds, because short clips leave less room to hide mistakes. For every clip we recorded:
- Cost in Animax credits. For models that aren't available in Animax we show the equivalent credit cost, so the numbers are comparable.
- Resolution and clip length.
- Generation time, from submission to a finished video.
- Quality score out of 10, for realism, prompt adherence, lip sync and sound, and overall usability.
All clips are on the Animax YouTube channel, so you can watch and judge for yourself. One caveat applies throughout: every model got one attempt per prompt, and any model can have a good or bad day. Treat each result as a data point, not a verdict for all time.
Test 1: Speech in a Rainstorm
Our May tests showed that lip sync on a calm, single speaker is largely solved. So we made it harder: two speakers taking turns, under an umbrella, in heavy rain and thunder.
The prompt:
Two friends huddle under a small umbrella on a busy neon-lit city street at night in heavy rain as thunder rumbles. One turns to the other and says: "Have you tried Animax?" The other grins and replies: "Yes, it's amazing!" Rain pattering on the umbrella, passing traffic, distant thunder.
Each model has to assign each line to the right person, keep their lips in sync over loud rain, and pronounce "Animax" correctly.
Gemini Omni 1.1 Flash — 80 credits | 720p | 32s gen | 4s clip | ⭐ 9/10
By far the best clip in this round. Two believable people share the umbrella on a neon-lit street, the first line is delivered naturally, and the clip cuts to a close-up for the grinning reply, exactly as an editor would. The rain on her face and hair looks real, the voices are clear above the weather, and it was also the fastest generation in the whole comparison.
HappyHorse 1.1 — 112 credits | 720p | 103s gen | 4s clip | ⭐ 4/10
HappyHorse builds a rich scene, with steam rising from the street and traffic passing, and it handles the move to a closer two-shot smoothly. But the people look noticeably artificial, and it never feels like real footage. It was also the most expensive clip in the round.
Sora 2 — 80 credits | 720p | 58s gen | 4s clip | ⭐ 3/10
The surprise of the test: Sora rendered the whole scene as anime, complete with Japanese street signs and a lightning flash. The prompt never asked for photorealism, so this is a legitimate interpretation, but it's a reminder that Sora takes stylistic liberties unless you spell out the look you want. The characters are small in the frame and the voices are weak.
Rainstorm verdict: Gemini Omni 1.1 Flash, by a wide margin. It's the only clip of the three you'd use as it is.
Test 2: The Instagram Reel
The second test is the kind of clip people actually post: a vertical 9:16 food video with a person in it. Faces, hands, a knife, juice and a laugh, all in four seconds, are hard to get right.
The prompt:
Instagram Reels style, handheld medium shot in a bright, sunny kitchen: a young woman with a big, warm smile slices a ripe watermelon in one quick stroke, juice splashing across the wooden board. She grabs a fresh wedge, takes a big crunchy bite and laughs at the camera, juice on her lips. Quick push-in from her hands to her face. Vibrant summer colours, natural window light, shallow depth of field, realistic hands and expressions. Audio: a crunchy chop, a juicy splash, her laugh, upbeat summer music.
Gemini Omni 1.1 Flash — 80 credits | 720p | 45s gen | 4s clip | ⭐ 8/10
The most realistic result of the five. Her face, smile and bite look genuinely natural, and the warm golden light suits the format. The chopping gets little screen time, and instead of a smooth push-in it jumps to a close-up, but the clip is the most convincing of the set.
HappyHorse 1.1 — 112 credits | 720p | 129s gen | 4s clip | ⭐ 7/10
The closest match to the prompt: one continuous shot that goes cut, wedge, bite, laugh, with the requested push-in from hands to face. Very lifelike overall. The final laugh is a little exaggerated, which slightly spoils an otherwise strong clip.
Veo 3.1 Lite — 40 credits | 720p | 168s gen | 4s clip | ⭐ 6/10
Veo produces the most natural-looking person of the five, but the story doesn't hold together: there's watermelon juice on her lips before she has eaten anything, and the board is already covered in cubes. It was also the slowest generation in the whole comparison.
Sora 2 — 80 credits | 720p | 58s gen | 4s clip | ⭐ 5/10
Clean and natural, with juice that looks like juice, but a wedge of watermelon appears from nowhere, which is easy to spot once you've seen it. The camera barely moves, so the push-in from the prompt is missing.
Seedance 1.5 Pro — 42 credits | 720p | 71s gen | 4s clip | ⭐ 3/10
Seedance follows the script step by step, but the watermelon juice is dark red and pools on the board like blood, and it runs from her mouth after the bite. In a food video that's a deal-breaker.
Reel verdict: Gemini Omni 1.1 Flash for realism, with HappyHorse a close second on prompt adherence. Comparing Seedance and Sora here makes a good lesson: the same script can come out appetising or unsettling.
Test 3: Bringing a Photo to Life
Image-to-video is a different skill: the model has to keep the subject from a real photo and then invent believable motion. We gave three models a still of a motocross rider carving down a dusty hill and asked for a dramatic camera move.
The prompt:
Start exactly from the input image: a motocross rider in red gear leaning hard into a turn on a steep, dusty hill. He twists the throttle and speeds forward down the slope, the rear wheel blasting a huge plume of dirt and dust. The camera performs a fast, smooth 180-degree orbit around the rider, sweeping from the front around his side, and finally settles directly behind him, chasing him as he races forward down the track. Super dynamic, high-energy action, realistic physics and motion blur, dust filling the air, bright natural sunlight. Audio: a roaring motocross engine revving hard, knobby tyres tearing through loose dirt, rushing wind.
HappyHorse 1.1 — 140 credits | 720p | 142s gen | 5s clip | ⭐ 8/10
The most realistic result by a distance. HappyHorse swings in close to the rider and then pulls out behind him as he races away in a cloud of dust, and it feels like footage from a real action camera. It also keeps the original photo's proportions instead of cropping it. The downsides are price and patience: the most expensive clip and the slowest of the three.
Gemini Omni 1.1 Flash — 100 credits | 720p | 39s gen | 5s clip | ⭐ 6/10
Gemini follows the camera direction almost literally, moving from the front to the side and then behind him in one continuous move, and it's very fast. The result is solid but doesn't quite reach HappyHorse's realism. It only outputs 16:9 or 9:16, so the photo gets cropped slightly.
Seedance 1.5 Pro — 50 credits | 720p | 77s gen | 5s clip | ⭐ 6/10
Seedance ends up behind the rider too, but by cutting between what look like separate shots rather than with one camera move, and the bike and kit change slightly along the way. It's close to Gemini in quality at half the price.
Photo-to-video verdict: HappyHorse 1.1. If you need the most realistic motion from a still, it's the one to beat. Gemini and Seedance are roughly level, and Seedance costs half as much.
Test 4: When the Leaderboards Get It Wrong
The simplest prompt in the test turned out to be the most revealing: one person, one line, no weather, no tricks.
The prompt:
A friendly woman in a bright modern office looks directly at the camera and says: "Welcome to our quarterly update." Soft office ambience in the background.
Wan 3.0 and MiniMax H3 sit near the top of the public leaderboards, so we expected them to breeze through this. They didn't.
Sora 2 — 80 credits | 720p | 77s gen | 4s clip | ⭐ 8/10
Clean, natural delivery with accurate lip sync, and it stops talking when the line ends. This is still the reference for a simple talking head.
Seedance 1.5 Pro Fast — 32 credits | 720p | 90s gen | 4s clip | ⭐ 6/10
The budget pick holds up well: the audio is clean and the line is delivered correctly. The visuals are a step below Sora's, but at 40% of the price it's a sensible everyday option.
Veo 3.1 Lite — 40 credits | 720p | 97s gen | 4s clip | ⭐ 4/10
Veo gives the speaker an accent the prompt never asked for, and her mouth keeps moving after the line has finished, the same overrun we saw in May.
Wan 3.0 — 80 credits | 720p | 77s gen | 4s clip | ⭐ 3/10
The picture is fine, but after "quarterly update" the speaker carries on in gibberish to fill the rest of the clip. For any video with a script, that rules it out.
MiniMax H3 — 64 credits | 768p | 71s gen | 4s clip | ⭐ 3/10
The same failure as Wan: good visuals, then gibberish once the scripted line runs out.
Talking-head verdict: Sora 2. Leaderboards measure what voters prefer when they watch two clips side by side, often without a strict script. If your video depends on someone saying exactly the right words and then stopping, test it yourself.
Bonus: Our May Prompts, Revisited
In May, Seedance failed both of our hardest scenes: the bar scene came out with the wrong characters, and the skydiver had a comically small parachute. We gave the same two prompts to Seedance 1.5 Pro Fast, the variant we now use for everyday clips.
Friends in a bar — 48 credits | 720p | 136s gen | 6s clip | ⭐ 5/10 (May: 2/10)
A clear step up from May. This time the cast matches the prompt, the story plays out as written, and the voice is fine. The details still let it down: the friends drink from coffee mugs in a bar, and the phone sometimes stands upright in a cup.
Skydiver — 40 credits | 720p | 90s gen | 5s clip | ⭐ 4/10 (May: 4/10)
The camera move and her excitement work, but the tiny parachute has been replaced by a cable running out of her helmet, so she looks like she's hanging from a wire rather than falling. Physics in high-motion scenes is still Seedance's weak spot.
Overall Results
Not every model took part in every test, so here is each model's average across the tests it was in:
| Model | Rainstorm | Reel | Photo-to-video | Talking head | Avg quality | Avg gen time | Credits per second (720p) |
|---|---|---|---|---|---|---|---|
| Gemini Omni 1.1 Flash | 9/10 | 8/10 | 6/10 | — | 7.7 | 39s | 20 |
| HappyHorse 1.1 | 4/10 | 7/10 | 8/10 | — | 6.3 | 125s | 28 |
| Sora 2 | 3/10 | 5/10 | — | 8/10 | 5.3 | 64s | 20 |
| Veo 3.1 Lite | — | 6/10 | — | 4/10 | 5.0 | 133s | 10 |
| Seedance 1.5 Pro / Pro Fast | — | 3/10 | 6/10 | 6/10 | 5.0 | 79s | 8–10 |
| Wan 3.0 | — | — | — | 3/10 | 3.0 | 77s | 20 |
| MiniMax H3 | — | — | — | 3/10 | 3.0 | 71s | 16 |
Gemini Omni 1.1 Flash is the standout of the new generation. It produced the most realistic people, the best speech in difficult conditions, and it did it faster than anything else, at the same price as Sora 2.
HappyHorse 1.1 has the highest ceiling for animating a photo, but it's the most expensive model here and inconsistent from text: superb in one test, unrealistic in another.
Sora 2 remains excellent at a clean, scripted talking head, but it now trails on realism, it can wander into its own visual style, and it's no longer the obvious premium choice. Its remaining advantage is length: up to 20 seconds per clip.
Veo 3.1 Lite still produces natural-looking people at a low price, but its logic and speech slips add up.
Seedance 1.5 is the value champion. It isn't the right model for complex or high-motion scenes, but for everyday clips it delivers at a fraction of the cost, and its bar scene was a big step up on May.
Wan 3.0 and MiniMax H3 show why public leaderboards are a starting point, not an answer: both rank near the top, and both failed the simplest scripted line.
What We Changed in Animax
We use these tests to decide which models power Animax, and this round led to a new line-up:
| Model | Best for | Credits per second (720p) | |
|---|---|---|---|
| Standard | Seedance 1.5 Pro Fast | Everyday clips; silent on the free plan, with sound on paid plans | 4 silent / 8 with audio |
| Gemini | Gemini Omni 1.1 Flash | The most realistic people and speech, 3–10 seconds | 20 |
| Sora | Sora 2 | Longer clips, up to 20 seconds | 20 |
Veo 3.1 Lite has been retired from the text-to-video line-up, and the Animax AI assistant now uses Gemini Omni whenever it adds an AI clip to a paid user's video. Image to video runs on Seedance 1.5 Pro Fast and Seedance 1.5 Pro, which matched Gemini on our motocross test at half the price. HappyHorse impressed us, but not by enough to justify its price for most users, so we've left it out for now.
About Animax.ai
The clips in this comparison were generated with the same models that power Animax.ai, and generative video is only one part of the platform.
Animax is an all-in-one video creation platform for content creators, marketers and businesses. Alongside AI video, it lets you:
- Build editable video scenes from plain language. Describe what you want, and adjust every detail afterwards.
- Edit every element. Change colours, timing, text, motion and style without wrestling with a timeline.
- Combine AI with your own assets. Mix generated clips with your brand assets, footage and audio.
- Create at scale. Produce ads, social content, product videos and presentations from a single workflow.
Try it at animax.ai.
Get Animax updates
New features, improvements, and fixes — delivered straight to your inbox. No spam, unsubscribe any time.