Which AI actually understands video?
68 models from 19 companies, tested on our proprietary data set. Ranked by how well they follow what happens on screen, and by what an hour of video costs.
Updated October 2026
Vote on the next benchmark ↓
81.1 ±1.9 / 100 · $6.86 per hour #1 of 68 69.1 ±4.2 / 100 · $1.38 per hour 12.0 points behind #1, 5× cheaper 
61.1 ±9.8 / 100 · $1.34 per hour #18 overall The leaderboard
Every model appears once, at the best setting we tested. Missing one? Request it.
81.1±1.9 $6.86 Oct 2026 
77.7±4.0 $46 Sep 2026 
77.5±3.1 $50 Oct 2026 
76.7±6.2 $77 Sep 2026 76.1±6.9 $3.31 Sep 2026 74.6±2.6 $23 Sep 2026 73.8±8.1 $110 Sep 2026 72.4±2.0 $16 Sep 2026 
69.4±2.6 $3.92 Feb 2026 69.1±4.2 $1.38 Aug 2026 68.9±5.7 $13 Aug 2026 67.2±6.2 $5.27 Jul 2026 66.7±4.3 $7.92 Aug 2026 66.0±4.6 $62 Jul 2026 65.4±10.4 $8.73 May 2026 64.7±3.6 $20 Sep 2026 61.5±12.2 $12 Feb 2026 61.1±9.8 $1.34 Aug 2026 59.0±11.5 $4.05* Feb 2026 58.9±5.2 $39 Jul 2026 58.9±5.7 $1.16 Feb 2026 58.6±5.7 $2.96 Aug 2026 57.4±3.6 $3.51 Jun 2026 
Z.ai 56.1±6.2 $1.02 Aug 2026 
52.3±8.3 $14 Jul 2026 52.1±5.2 $6.08 Aug 2026 51.7±1.9 $1.24 Oct 2026 50.3±4.3 $3.67 Dec 2025 48.3±3.1 $113 Mar 2026 47.9±3.6 $14 Sep 2026 47.4±4.5 $2.29 Feb 2026 46.6±3.5 $2.19 Apr 2026 46.4±8.4 $19 Jun 2026 
45.7±9.6 $0.96 Sep 2026 45.2±2.9 $10 Mar 2026 45.1±4.2 $1.23 Jul 2026 44.5±7.9 $1.40 Mar 2026 44.5±5.6 $15 Aug 2026 44.0±6.4 $3.97 Feb 2026 43.7±2.8 $5.20 Apr 2026 43.6±6.7 $0.75 Sep 2026 42.8±7.7 $7.47 Jul 2026 42.8±3.6 $1.85 Apr 2026 41.6±3.4 $8.54 Apr 2026 40.8±6.3 $14 Sep 2026 40.7±3.7 $7.65 May 2026 
39.9±11.4 $1.95 Sep 2026 37.9±6.2 $0.81 Sep 2026 36.7±2.2 $0.74 Jul 2026 35.7±8.4 $3.12 Sep 2026 35.3±4.3 $0.47 Apr 2026 
35.3±3.6 $3.36* Oct 2026 
35.1±3.8 $5.30 Oct 2026 29.5±7.3 $0.37* Jun 2026 27.9±6.4 $0.38 Mar 2026 
26.8±3.3 $0.34 Sep 2026 24.5±3.0 $8.03 Oct 2025 23.2±3.3 – Apr 2026 21.5±3.9 $2.56* Apr 2026 19.2±3.3 $3.08* Feb 2025 
13.7±4.7 – Aug 2026 
9.7±5.5 – Aug 2026 
5.4±2.1 $0.04 Mar 2026 No models match. Request it
Score is out of 100; higher is better. ± is the 95% confidence interval across five runs of each model (four for three of them), drawn as the bracket on each bar. Where two models' ranges overlap, the gap between them is within run-to-run noise.
$ / hour is what it costs to describe an hour of footage in enough detail to know what is happening at any moment. It is not the price of sending a whole hour of video in one request. Prices are OpenRouter's per-token rates on the day we tested each model, with reasoning tokens counted as output. A dash means we have no price for that run.
* Seed 2.0 Pro: not on OpenRouter, so priced at BytePlus's own per-token rate. Pegasus 1.6, Pegasus 1.5 and Pegasus 1.2: not on OpenRouter, so priced at TwelveLabs' own rate: per minute of video, plus output tokens. MiniCPM-V 4.7 35B: no OpenRouter listing, so priced at what it costs us to run it on a rented H200 GPU ($5.81 an hour). DiffusionGemma 26B: no OpenRouter listing of its own, so priced as Gemma 4 26B.
The new #1, at a fraction of the price
Gemini 4 Argon scores 81.1, 3.4 points ahead of Claude Opus 5.5, at $6.86 an hour instead of $46. Cheaper still, Gemini 3.7 Flash scores 69.1 at $1.38.
Seeing a frame is easy. Following a video is hard.
Describing a single image is close to a solved problem. Video is different: the meaning is in what changes. Who does what, in what order, and how it ends.
Most models never actually watch the video. They're handed about one frame per second and have to reconstruct the action from stills. Anything that happens between two frames, they never see.
Plenty of the models here were never sold as video models, and never claimed to be. We tested them anyway, because many of them can piece together what happened from a handful of frames. Some of the best results on this board come from exactly those models.
Understanding what's actually happening in video is the problem we work on every day at StarZero. We built this leaderboard to know which models to trust, and we'll keep it updated as new ones ship.
Missing a model? Request it.
Send us a link to it: its Hugging Face page, its OpenRouter listing, or the maker's API docs if it's private. We'll look at adding it to the board.
Leave your email and we'll tell you when it's on.
Help shape the next benchmark.
The data set behind this board is proprietary, and it stays that way. No one can train on it or tune for it.
But models have caught up fast over the past few months, so we're thinking about a new, harder one. Should it stay proprietary too, or would you rather have an open-source version, instead or as well? Tell us, and subscribe to hear when the board changes.
Talk to your whole video library.
StarZero's agents search it, cut it, caption it, reframe it, and render finished clips for every platform, all from one chat.