VLMs on video

VLMs on video: price vs performance

Gemini 3.7 Flash nearly matches the top tracking score at $4.86 per video hour, against $44 for Claude Opus 5.5 and $105 for GPT-6 Astra.

Updated 25 Sep 2026 · Perception Test · same pipeline for every model · 300 train questions

Tracking quality and cost

ModelHOTA, pipeline$ / video hourRight object
Claude Opus 5.50.572$4492%
Gemini 3.7 Flash0.569$4.8690%
GPT-6 Astra0.567$10597%
Gemini 3.8 Flash0.565$8.73not scored
Qwen 3.8 27B0.476$2.43*82%

300 grounded video QA questions from DeepMind's Perception Test (train split). Cost is the API bill per hour of video. *Qwen is self-hosted on Agnify; $2.43 is Agnify pricing.

GPT-6 Astra picks the right object most often (97%), but ties on tracking.

Methodology and limitations

What each model can do with video

ModelVideo inputBoxesMasksVision Evals (images)
Gemini 3.7 / 3.8 FlashNative videoYesYes, polygons85.2%
Claude Opus 5.5Images onlyYesUndocumented prompting, polygons85.5%
GPT-6 AstraImages onlyYesUndocumented prompting, polygons86.6%
Qwen 3.8 27BNative videoYesUndocumented prompting, polygons74.7%

Vision Evals is an external image benchmark, averaged over six tasks (Gemini 3.8 Flash: 85.1%). Our pipeline uses SAM 3 to track the selected object across frames.

Cost calculator

Custom workload

Input Output

How the cost is calculated

Costs are API list prices for the tokens each model reads and writes, without tracking GPU time. Qwen is Agnify pricing.

How frames turn into tokens

Every model bills by tokens, and each video frame becomes a fixed number of tokens before the model sees it. Gemini takes the clip as video and billed us a flat 66 tokens per frame; Google's table for Gemini 3 lists 70 at the default resolution and 280 at high (Gemini docs). Claude and GPT-6 take no video, so every frame is sent as an image. Claude bills one token per 28x28 px patch, so a 512x288 frame costs about 210 tokens and a full-HD still 2,691 (Claude docs). GPT-6 Astra comes to about 180 tokens per 512 px frame on our bills, in line with independent measurements. The model's answer, including its reasoning, bills at the higher output rate, and none of these models lets you turn reasoning off. Qwen runs on our own GPUs at Agnify pricing.