Claude 從螢幕錄影中實際接收甚麼:4 種傳送方式的 token 數量
What Claude actually gets from a screen recording, token counts for 4 ways to send it
作者以同一段 43 秒、涵蓋 3 個畫面的 X.com 操作錄影,比較 4 種交給模型的方式及 token 開銷;Claude 不能直接開啟錄影,須先轉成可讀輸入。
Claude can't open your screen recording. Anthropic's vision docs list JPEG, PNG, GIF and WebP as input formats, and for a GIF only the first frame is used. Codex and ChatGPT take images too, not MP4.
So when a bug needs a 43 second repro, something has to turn the video into what the model can read. I tried four ways on the same recording: 43 seconds, 3 screens, a few clicks on X.com.
1. Still frames
Each frame is an image, and Claude 4.7 counts image tokens in 28px patches: ceil(width/28) x ceil(height/28). A 1440x900 frame is 52 x 33 = 1,716 tokens.
One frame per screen is 3 frames, about 5,150 tokens. One frame per second is 43 frames, about 73,800 tokens, and once a request carries more than 20 images the per-image size limit gets stricter.
Either way the clicks are gone. You see the screen before and after, not which button was pressed.
2. One tall stitched page
Stitching the whole flow into one long image looks clever and costs almost nothing. That's the problem.
Claude 4.7 shrinks any image to 2576px on the long edge (older models 1568px). A 2540x18410 page arrives as 355x2576, about 1,196 tokens, and 16px body text ends up around 2.2px tall. On older models it's 216x1568 and 1.4px.
The same thing happens with Playwright fullPage: true captures. Odilon Hugonnot's CLAUDE.md for mobile redesign uses full-page screenshot scripts at 375 to 768px, and on a long mobile page those get squeezed the same way before Claude checks them. Crop to the viewport, or slice the page into tiles with a small overlap so a line cut at the seam still shows whole in one tile.
3. Gemini, which does take video
Gemini is the one model here that accepts video. It samples 1 frame per second at about 100 tokens per second at default resolution, so the 43 second clip is about 4,300 tokens.
Cheapest of the four. It still only sees one frame a second, so a quick click between two samples can disappear, and you're outside Claude or Codex for that question.
4. The recording as JSON
This one skips pixels. SlimSnap records the screen with voice and writes JSON next to the MP4: one frame per screen with its text and positions, every click with the control it hit and the screen it led to, and each spoken sentence tied to the element under the pointer.
The same 43 seconds came to 6.3k tokens of JSON.
It costs more than Gemini and far less than a frame per second. What you get for it is that "I clicked Save and nothing happened" reaches the model as the actual control plus the screen that came after.
Which one to send
| You want to show | Send |
|---|---|
| One screen | One cropped screenshot |
| A long page | Viewport-sized tiles, not one tall image |
| A flow, and you use Gemini | The MP4 |
| A flow, and you use Claude or Codex | One frame per screen, or JSON with the clicks |
Reproduce the numbers
Tokens for one image on Claude 4.7:
python3 -c "import math;w,h=1440,900;print(math.ceil(w/28)*math.ceil(h/28))"
The size Claude actually sees for a tall page:
python3 -c "w,h=2540,18410;s=2576/max(w,h);print(round(w*s),round(h*s))"
A still at each screen change, to send as frames:
ffmpeg -ss 4.5 -i repro.mp4 -frames:v 1 screen2.png
來源:DEV Community · dev.to