All benchmarks

VGI-Bench

VGI-Bench (Video General Intelligence Bench, by Seldon) asks multiple-choice questions about long-form videos: what appeared first, what was on screen at the same time as something else, what a narrator said while a specific shot was up, and questions whose premise never happens. Each sample below is a real run through OpenRouter. The model received the full video and the question in one turn at temperature 0, and the graded answer is the letter on its final Answer: line.

Last benchmark run

Model comparison

Most Accurate

Best Value

Accuracy
Representative-run accuracy, best first.
Cost per question
Average cost per question, cheapest first.
Time per question
Average wall-clock time per question, fastest first.

Cost efficiency

Accuracy vs. cost (Pareto frontier)
One point per model, using default routing (not pinned to a provider) when available. The line is the Pareto frontier: no model beats these on both accuracy and cost.

Leaderboard

Accuracy over the public split (439 questions, runs need at least 395 answered to be listed), with cost, time, and input and output tokens per question for each provider endpoint the model was run on. Scores aggregate successful runs from the newest 90 days.

VGI-Bench leaderboard: model accuracy, cost, time, and input and output tokens per question by provider
#ModelStd dev
166.2%±1.0pp$0.03964s141k2.03k
265.5%--$0.0464.6m412k11.5k
363.7%±4.8pp$0.0732.5m241k6.5k
463.7%--$0.558.6m613k20.1k
563.6%--$1.033.1m233k8.62k
6
Google AI Studio
61.4%±0.7pp$0.0611.8m3416.04k
761.1%--$0.224.2m227k14.3k
858.0%--$0.338.2m579k13.6k
957.4%--$0.0231.6m231k9.18k
1056.8%----1.9m313102
1156.8%--$0.116.2m1.52M24.1k
1254.8%±0.5pp$0.144.7m111k40.2k
1354.8%±0.5pp$0.275.6m1.48M10.9k
1454.2%±0.5pp$0.242.6m151k21.8k
1552.8%±2.3pp$0.163.7m98.6k14.1k
1651.3%--$0.0086.0m7.53k13.7k
1751.0%±0.6pp$0.162.0m150k15.3k
1850.6%±0.4pp$0.217.1m532k20.9k
19
Google AI Studio
50.0%±0.3pp$0.0241.9m3419.42k
2048.0%--$0.0122.4m34029.7k
2147.6%±1.9pp$0.0262.3m36.7k7.15k
2246.5%±2.9pp$0.0186.4m272k13.2k
2346.2%±0.1pp$0.0082.7m46.4k817
2446.2%±1.2pp$0.104.9m1.49M20.2k
2545.6%--$0.0274.6m77.4k14.7k
2645.5%±1.0pp$0.958.6m1.45M21.4k
2745.5%±0.7pp$0.177.5m532k21.3k
2845.4%--$0.0345.2m337k11.9k
2945.2%±0.8pp$1.146.7m1.49M19.4k
3045.1%±0.6pp$0.174.9m440k18.6k
3143.4%±2.9pp$0.0271.9m232k9.34k
3243.4%--$0.343.2m77k21.1k
3342.3%--$0.193.5m515k7.53k
3440.9%±7.2pp$0.704.4m38.3k81k
3534.8%±2.9pp$0.0217.5m199k16.4k
3634.6%±2.7pp$0.0288.3m58.2k43.9k
3730.1%±2.6pp$0.0819.8m53.4k75.3k
3828.9%--$0.0191.6m49k7.48k
3928.5%±0.8pp$0.0161.9m14.6k30.3k
4024.7%±7.3pp$0.0042.8m4.59k9.91k
4122.2%--$0.00644s130k3.99k
42
Reka Edge
Pareto
Reka
21.4%±0.2pp$0.0001531s1.49k30.5
4319.2%--$0.0011.8m5.71k9.46k
4415.5%--$0.05044s1.58k4.7k
4514.7%--$0.323.0m1.63k15.5k
4614.4%--$0.03976s1.42k10.3k
4713.8%--$0.02247s6.28k6.46k
4812.8%±0.4pp$0.0233.9m38k6.22k
4912.7%±0.0pp$0.06851s1.58k8.24k
5012.4%--$0.01324s1.19k3.29k
5112.4%±0.4pp$0.01556s38.3k6.56k
5212.3%±0.5pp$0.136.9m124k2.42k
5311.8%±0.4pp$0.0313.9m32.7k5.87k
5411.8%--$0.04182s5.06k3.17k
55
Google AI Studio
11.5%--$0.01624s1.39k3.97k
5611.4%--$0.111.9m1.58k10.6k
5710.9%--$0.09859s1.39k7.96k
5810.8%±0.3pp$0.00544s4.57k5.08k
5910.6%--$0.09860s1.39k8.02k
60
Google AI Studio
10.1%--$0.08839s1.38k9.59k
618.4%--$0.00613s1.39k2.08k
628.2%--$0.03858s1.39k25.1k
638.2%--$0.03858s1.38k25.1k
646.4%--$0.2657s1.57k12.9k
656.3%±5.7pp$0.02536s5.46k1.69k
665.6%±4.3pp$0.0171.9m69.2k804
673.3%±1.5pp$0.0037.4m61221.4k

Trajectories

The reasoning trace is the model's perception: the moments it says it saw and the timestamps it attaches to them. Reading it against the video shows how the answer was reached, and on the failed runs, where the description parts from the footage.

correct

Which window the pointer hovers over most

Happy path: The model locates the moment and reads it correctly.

Question 1187 global-graph/interaction_partner Qwen3.8 27B

The full video the model was sent, served from the OpenRouter mirror used by the harness.

Which option does the mouse pointer hover over most often?

  1. A)Windhawk application windowkey, model
  2. B)Install button
  3. C)Adobe Premiere Pro window
  4. D)Windows taskbar
answered A ✓key A

The trace tracks the pointer across the whole recording rather than one frame, names the Premiere Pro cameo near the end as a distractor, and lands on the window that dominates the session.

The question asks which option the mouse pointer hovers over most often. Let me think about the video content. The video is a tutorial about Windhawk, a customization tool for Windows. Throughout the video, the user is constantly interacting with the Windhawk application window - browsing mods, installing them, configuring settings, etc. Let me consider each option: - A) Windhawk application window - This is the main focus of the video. The user opens Windhawk, browses mods, clicks on details, settings, installs mods, etc. The mouse is frequently over this window. - B) Install button - The install button is clicked a few times but not hovered over most often. - C) Adobe Premiere Pro window - This appears only briefly near the end (around 13:20-13:40) as a joke about the title bar color. - D) Windows taskbar - The taskbar is shown but the mouse isn't constantly hovering over it. The dominant interaction is with the Windhawk application window. The mouse pointer is over it for the majority of the video as the user navigates through mods, settings, etc. So the answer is A. Let me double check by reviewing timestamps. From 1:20 onwards, the Windhawk window is open and the mouse moves around it constantly - clicking Mods, Explore, scrolling, clicking Details, Settings, etc. This continues through most of the video. The taskbar is shown at times but the mouse hovers over the Windhawk window more. Thus A is correct.