Tool 03 · context
What 256K of context really holds
Convert abstract token counts into pages, frames and video minutes, then compare 24 models under one sampling policy.
256K · 80% usable context
256
document pages
682
video frames
11
min video @ 1fps
A theoretical ceiling, not a quality ceiling. Beyond ~60% occupancy, recall of middle details often drops; pair with chunked summaries or retrieval.
What each model holds under the same settings
800 tokens/page · 300/frame · 1fps · 20% reserve
p = pages, m = video minutes. Models with undisclosed context count as 0 and sort last.
About context
Does longer context always help?
No. More frames dilute attention and positional extrapolation drifts. Long-video quality depends more on sampling, hierarchical summaries and retrieval than raw context.
Why reserve 20%?
Prompts, system instructions and output all consume tokens. Filling the window squeezes output, latency and cost.
Why do frame tokens vary so much?
Vision encoders, tiling and pooling differ by vendor. 300 tokens/frame is a median assumption; the real range can be 80–800.
Personal ranking · Not an official score
Every ranking, score and comparison on this page is an independent personal ranking, not an official score, and represents no vendor, institution or regulator. Scores are a subjective mapping of public evidence under our own weights — change the weights and the order changes. Disagree, or spotted a factual error? Contact the site owner: jizhi2024@foxmail.com — we check each report and log corrections in the changelog.