Why image size matters
Vision models convert images into patches or tiles, so token cost grows with image area. Halving both dimensions usually cuts token cost to about a quarter.
Vision token estimates
Upload an image or enter its dimensions to estimate how many tokens it costs on GPT-6, Claude, and Gemini — plus what a single image costs at current API prices.
Estimates from each provider's documented formula and current input pricing.
| Model | Tokens | Est. cost / image |
|---|---|---|
| GPT-6 AstraOpenAI | 1,229 | $0.012 |
| GPT-5.6 SolOpenAI | 1,229 | $0.0049 |
| GPT-5.6 LunaOpenAI | 1,229 | $0.0002 |
| Claude Sonnet 5Anthropic | 1,369 | $0.0027 |
| Claude Haiku 4.5Anthropic | 1,369 | $0.0014 |
| Gemini 3.8 FlashGoogle | 1,032 | $0.0008 |
| Gemini 3.5 Flash-LiteGoogle | 1,032 | $0.0003 |
Vision models convert images into patches or tiles, so token cost grows with image area. Halving both dimensions usually cuts token cost to about a quarter.
For most analysis tasks, 768 to 1024px on the long edge is plenty. Anything larger mostly buys token cost, not accuracy.
Multiply the per-image cost by your expected volume. A thousand phone photos can cost several dollars on premium models, but cents on lite ones.
It depends on the model and the image size. A 1024x1024 image is roughly 765 tokens on tile-based models, about 1,600 tokens on Claude's patch counting, and 258 to 1,032 tokens on Gemini depending on tiling. Larger images cost more because cost grows with image area.
Current OpenAI models use 32px patches: tokens are roughly 1.2 times the patch count, capped at 1,536 patches. Older GPT-4o and GPT-4.1 used 512px tiles at 170 tokens each plus an 85-token base, or a flat 85 tokens in low-detail mode.
Claude divides the image into 28x28 pixel patches: tokens equal ceil(width/28) times ceil(height/28). Images whose longest edge exceeds 1568px are downscaled by the API before counting.
Gemini charges a flat 258 tokens for images up to 384x384 pixels. Larger images are split into tiles, with 258 tokens per tile, so cost steps up as the image grows.
Resize images to the smallest size that still shows what the model needs, usually 768 to 1024px on the long edge. Halving both dimensions cuts token cost to roughly a quarter, and lite models charge far less per token for vision.
No. This calculator reads only the image width and height locally in your browser. The image file itself is never uploaded anywhere.