What we learned benchmarking 11 video models on GMI Cloud

I sent three prompts to eleven text-to-video models on GMI Cloud, at 720p and five seconds, through the same queue, on the same afternoon. I timed every clip from the server side, priced every clip with the estimate endpoint, scored every clip against a rubric, and then ran image-to-video and resolution sweeps on top. Forty-five clips, $35.86 at list price. The fastest model returned a five-second clip in 32 seconds and the slowest took five minutes. The cheapest model per second scored the same as the most expensive.

Why I ran this

Video is where the price tags get serious. A five-second clip costs between $0.22 and $2.00 on this catalog, and a 30-second 1080P clip on Wan 3.0 is $6. The public arenas rank these models by human preference, but they do not say what a clip costs on GMI, how long you wait, whether the audio track is real, or whether the model will take your first frame. I wanted those numbers in one table, measured the same way for every model.

The Models

The eleven models, with the settings I used and the list price the estimate endpoint quoted for exactly those settings:

Veo only offers 4, 6 or 8 seconds, so I used 6, its default. Every other model got 5. Cost figures below are normalized to a price per second so the Veo clips compare fairly.

How I measured

Cost is the list price from the estimate endpoint for the exact parameters submitted. The estimate can differ from the catalog card because the card shows a base rate: Seedance 2.0 at 720p for five seconds quoted $0.76, which is $0.152 per second against a $0.07 card rate. Estimate with the parameters you will send.

Time is the difference between the queue's created_at and the Last-Modified header on the mp4 in storage. Both are server clocks. I submitted all eleven requests for a prompt within about 35 seconds and polled, so the models ran concurrently, the way a real batch would.

Quality is a 0 to 5 score per prompt from one rater, me, from the clips and from first, middle and last frames laid out side by side. I also recorded a motion score per clip (mean absolute pixel difference between eight evenly spaced frames) and the mean loudness of the audio track, so a claim like "the audio is real" has a number behind it. One rater means a tenth of a point between two models is not a ranking.

The three prompts:

1. Barista. A barista pours steamed milk into a cup, latte art forms, steam rises, slow push-in, 35mm film look. Tests hands, fluid and a camera move.

2. Coastal drone. A low drone shot forward along a coastal road at golden hour, waves on rocks to the left, a red vintage convertible driving toward the horizon. Tests camera motion, a moving subject and a left/right instruction.

3. Retriever. A golden retriever climbs out of a lake onto a dock and shakes the water off, droplets in slow motion, handheld, overcast. Tests animal motion and water.

A note on the numbers. Every number in this post comes from the GMI Cloud MaaS inference engine on a Tier 5 account, so your cost and wait times will vary with your account tier, queue load on the day, and the provider's own capacity. If you need faster or more predictable turnaround, a dedicated endpoint runs your model on reserved capacity and cuts the queue wait out of the picture.

The Cost Picture

estimate_generation on GMI Cloud, Oct 14 2026 · MiniMax at 768P, Veo normalized from 6 s clips

That is a 9x spread from PixVerse to Veo 3.1. Resolution and duration multiply it. Wan 3.0 is $0.10 per second at 720P and $0.20 at 1080P, so a 10-second 1080P clip is $2.00. Seedance 2.5 at 1080p is $2.845 for five seconds. Veo 3.1 Fast at 1080p for eight seconds is $1.20. PixVerse at 1080p with its audio switch on is $0.575, more than double its 720p price.

The Speed Picture

queue created_at vs storage Last-Modified, 33 text-to-video runs

Gemini Omni 1.1 Flash returned all three clips in exactly 32 seconds. Wan 3.0 returned all three in exactly 137 seconds. Those two have the tightest schedules in the set. Seedance 2.5 is the outlier at the slow end: 200 seconds on the first prompt, 417 on the second. If you need a clip inside a minute, the choices are Omni 1.1 Flash, Veo 3.1 Lite, PixVerse and Luma. Veo 3.1 Fast's median was longer than Veo 3.1's on this queue.

The sweeps follow the same pattern. The 10-second 1080P Wan 3.0 clip took 258 seconds. The 1080p Seedance 2.5 clip took 96 seconds, faster than its own 720p runs, which suggests most of the 720p time was queue wait rather than generation.

Test 1: Barista

Every model poured milk into a cup and most of them finished on a rosetta. The differences were in the camera. Seedance 2.0 and Seedance 2.5 committed to the push-in and ended on a macro of the finished art. Gemini Omni 1.1 Flash framed tight from the first frame. Veo 3.1 Fast added a beat where the barista offers the cup to camera. MiniMax H3 kept the hands and cup steady with faint steam. HappyHorse 1.1 stayed wide and skipped the push-in. Luma Ray 3.2 was dark and heavy with steam and lost detail in the cup. PixVerse was clean but silent at 720p.

Test 2: Coastal drone

This prompt separated the field more than the other two. Gemini Omni 1.1 Flash did everything asked: waves on the left, the car ahead on the road, a smooth forward track, golden light. Veo 3.1 Fast, PixVerse V6, Luma Ray 3.2 and MiniMax H3 were close behind. Seedance 2.0 cut the car off at the frame edge in the first frame and then lost it entirely. Seedance 2.5 put the car on a wet road between cliffs with no waves in sight. Wan 3.0 pointed the car at the camera instead of away from it. Veo 3.1 returned the clip letterboxed with black bars top and bottom, the only output in the set that was not full-frame 16:9.

Test 3: Retriever

Seedance 2.5 was the best clip of the day. The dog climbs a dock ladder, plants its feet and shakes with a visible spray of droplets. Seedance 2.0, Veo 3.1, Veo 3.1 Fast, PixVerse and MiniMax all got a real shake. Wan 3.0 and Omni 1.1 Flash got the dog out of the water and onto the dock but the shake was mild. Luma's shake was blurry. HappyHorse's was short.

Price against quality

Mean rubric score over the three prompts, with price per second:

The three models at the top are within a tenth of a point of each other, which from a single rater reads as about the same. They cost $0.08, $0.10 and $0.15 per second. The $0.40 model sat a sixth of a point below them, level with PixVerse at $0.045, which would be my pick for silent b-roll. On these prompts the $0.08 to $0.15 band is where the value is, and MiniMax H3 at $0.40 a clip is the strongest price-to-score ratio in the set with an audio track.

Image-to-video

I generated a first frame with GPT Image 2.5 Flare, a barista mid-pour at 1536x1024, and handed its public URL to every model with a first-frame parameter. The parameter names differ per model and are listed in data/models.json: first_frame on Wan, Seedance and HappyHorse, first_frame_image on MiniMax, start_frame_url on Luma, image on Veo, reference_image with task image_to_video on Omni.

Seven models animated it: Wan 3.0, Omni 1.1 Flash, Seedance 2.0, MiniMax H3, HappyHorse 1.1, Veo 3.1 Fast and Luma Ray 3.2. Every one of them kept the barista, the apron, the counter and the light, and five of the seven ended with the cup slid toward the camera as the prompt asked. HappyHorse and Wan pushed all the way into a macro of the latte art. Luma kept the wide shot and moved the least.

Output sizes also changed. Wan, Seedance 2.0, MiniMax, Luma and HappyHorse adapted to the 3:2 source and returned 1174x782, 1112x834, 1152x768, 1104x832 and 1176x784. Omni and Veo kept 1280x720.

What the parameters did

• Audio off (Seedance 2.0). generate_audio false produced a silent file. The flag works.

• Audio on (PixVerse V6). generate_audio_switch true at 1080p produced an AAC track at a mean of -38 dB, which is quiet background sound rather than speech or effects. PixVerse at 720p with the switch off is silent.

• 1080P and 10 seconds (Wan 3.0). $2.00, 258 seconds, 1920x1080, and the pour reads better at the longer length because the latte art has time to form.

• 1080p (Seedance 2.5). $2.845 for five seconds. The clip is sharper than the 720p run and arrived faster.

• 1080p and 8 seconds (Veo 3.1 Fast). $1.20. Clean, with the barista framed in a wider cafe.

• Audio tracks in general. Nine of eleven models return an AAC track by default. Luma and PixVerse at 720p do not. Mean loudness ranged from -15 dB on Wan 3.0's coastal clip to -41 dB on MiniMax's retriever, so "has audio" covers everything from waves and engine noise to near silence.

What the public leaderboards say

Artificial Analysis AA-Video-T2V v2.1 pins Kling 3.0 1080p Pro at 1000 Elo. Their text-to-video and image-to-video boards against my three-prompt mean:

LMArena's text-to-video board puts Gemini Omni 1.1 Flash first at 1213, with Seedance 2.0, Wan 3.0, Seedance 2.5 and MiniMax H3 within 45 points. FilmBench, a film-grade benchmark published in July, scores Seedance 2.0 at 88.9 and HappyHorse 1.1 at 87.4 on text-to-video.

My three prompts agree with the arenas on MiniMax H3, Omni 1.1 Flash and Seedance 2.5 near the top. They disagree on two models. Wan 3.0 leads the Artificial Analysis board and scored 4.0 with me, because it put the car facing the wrong way and barely shook the dog. The Veo family sits at the bottom of that board and scored 4.3 to 4.5 with me, because it followed every camera and subject instruction I gave it. Three prompts is a small sample, and arena voters see thousands. But it is a reminder that an Elo is an average over prompts you did not write.

Key Takeaways

1. The value band is $0.08 to $0.15 per second. MiniMax H3, Gemini Omni 1.1 Flash and Veo 3.1 Fast scored about the same at the top and all sit in it. Veo 3.1 at $0.40 did not score higher.

2. Omni 1.1 Flash is the latency pick. 32 seconds for every clip, text-to-video and image-to-video alike.

3. Seedance 2.5 is the quality pick when you can wait. Best single clip of the day and the slowest median.

4. PixVerse is the silent b-roll pick. $0.045 per second, 4.3 mean, no audio at 720p.

5. Normalize to price per second and resolution before you compare. Veo clips are 6 seconds, MiniMax is 768P, Seedance 2.0 estimates at double its card rate.

6. Image-to-video parameter names are per model. Seven models, five different names for the same input. Take them from the model's parameter list.

7. Audio flags work per model. Seedance 2.0's off switch produced a silent file; PixVerse's on switch at 1080p added a quiet background track.

Quick Reference

Try It on GMI Cloud

curl -s https://console.gmicloud.ai/api/v1/ie/requestqueue/apikey/requests \
  -H "Authorization: Bearer $GMI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gemini-omni-1.1-flash-preview",
       "payload": {"prompt": "A golden retriever climbs out of a lake onto a wooden dock and shakes the water off its fur, droplets flying in slow motion, handheld camera, overcast daylight, natural sound",
                   "task": "text_to_video", "durationSeconds": "5", "aspectRatio": "16:9", "resolution": "720p"}}'

# poll until status is success, then download outcome.media_urls[0].url
curl -s https://console.gmicloud.ai/api/v1/ie/requestqueue/apikey/requests/$REQUEST_ID \
  -H "Authorization: Bearer $GMI_API_KEY"

The Question to Answer

How long can your product wait for a clip, and does it need sound? If the answer is under a minute, you have four models and the best of them costs $0.50. If the answer is "whatever it takes," Seedance 2.5 will give you the best shake of a wet dog you have seen from a prompt, five minutes later. If the answer is "it has to start from this photo," check the parameter name first, because every model spells it differently.

Ready to build? Get Started