Vision Image Token Calculator and Resize Budget

How it works

Presets are dated 2026-11-14. GPT-4o, GPT-4.1 and GPT-4o-mini use base-plus-tile accounting. High/auto detail fits within 2048 pixels, then reduces a shortest side above 768 pixels. Coverage uses 512-pixel tiles; low detail uses only the base. GPT-4.1-mini (2025-04-14 snapshot) uses 32-pixel patches, a 2048-pixel dimension limit, a 6,144-patch budget and a 1.62 multiplier for all three offered detail levels.

tile tokens = base + ceil(width / 512) × ceil(height / 512) × tile rate patch tokens = ceil(patch coverage × multiplier) resize scale = max(minimum width / width, minimum height / height) cost = tokens per image × image count × input rate / 1,000,000

Original image mode ignores minimum dimensions. In Compare resize mode, integer cross-products select the controlling minimum exactly and integer ceiling division rounds the other dimension up; provider processing dimensions round down. Coverage and patch tokens round up. Intermediate ratios stay unrounded. Select a model’s override-rules option to show only its tile or patch controls and replace the preset accounting coefficients and resizing caps; custom results are labelled. No image file is changed. The editable $1 rate is an example assumption, not provider pricing. The tool accepts dimensions up to 1,000,000 pixels and counts up to 1,000,000 as local arithmetic limits.

Worked examples

Square tile resize

GPT-4o at high detail uses 4 tiles for a 1024 × 1024 upload: 85 + 4 × 170 = 765 tokens. A 512 × 512 minimum-size upload uses 255 tokens, saving 510. At an example $1 per million input tokens, one resized image costs $0.000255.

Patch accounting

A 1024 × 1024 GPT-4.1-mini image uses 32 × 32 = 1,024 patches. Rounding 1,024 × 1.62 up gives 1,659 tokens. Two such images use 3,318 tokens.

Frequently asked questions

Sources

• OpenAI vision — sizing behavior and image token accounting (checked 2026-11-14)