Prompt & Done · TILES

Claude reads your image as tiles. Here's what that costs, and how to get it read right

Claude doesn't look at a picture the way you do. It cuts the image into 28×28-pixel patches ("visual tokens") and reads those, like words.

Checked 7 Nov 2026 against Anthropic's vision docs and our own test in Claude Code 2.1.319.

1 The token formula

visual tokens = ceil(width / 28) × ceil(height / 28)

Example: an 800×500 image = 29 × 18 = 522 tokens. A 1000×1000 image = 1,296 tokens.

Too big? It gets shrunk first. Claude 4.7 and later (high-resolution tier): long edge up to 2,576 px and up to 4,784 visual tokens. Other models: 1,568 px and 1,568 tokens. A 4K screenshot (3840×2160) is downscaled to 2576×1449 (4,784 tokens) on the high-res tier. Hard limit: 8000×8000 px; with more than 20 images in one request keep each under 3,000 px.

platform.claude.com: Vision → Resolution and token cost

2 The crop tip

When the answer is in small text (a code, an error line, a number in a table), crop to that region before you paste. Shrinking a huge screenshot blurs small text; a crop keeps it sharp and costs fewer tokens.

Our test: the same dashboard screenshot, same question ("What is the refund approval code?"). Full 6000 px image: Claude Code sent it as 1999×1133, 3,136 input tokens, answer "UNREADABLE". The 800×500 crop: 661 tokens, answer "QX7-48K2", correct. One synthetic screenshot, one model; your results will vary.

Don't over-crop: keep the context Claude needs to understand what it's looking at (Anthropic's own advice).

3 One-line check before you paste

On a Mac, this prints an image's size and its approximate visual tokens (before any downscaling):

f=shot.png; w=$(sips -g pixelWidth "$f" | awk '/pixelWidth/{print $2}'); h=$(sips -g pixelHeight "$f" | awk '/pixelHeight/{print $2}'); echo "$w x $h -> $(( (w+27)/28 * ((h+27)/28) )) tokens"

Over ~2,576 px on the long edge? It will be shrunk. If you need small text read, crop first.