IMAGE TOKEN CALCULATOR
Calculate vision tokens for GPT, Claude, and Gemini.
Images are billed in tokens just like text, but each provider counts them differently. Enter your image dimensions to see the token cost per provider rule set and the estimated price per image at current input rates.
Image tokens
765
Estimated input cost per image
Calculating...
Common image sizes as tokens
| Resolution | OpenAI (high detail) | Claude | Gemini |
|---|---|---|---|
| 512 × 512 (icon / thumbnail) | 255 | 350 | 258 |
| 768 × 768 | 765 | 787 | 258 |
| 1024 × 1024 (AI-generated square) | 765 | 1,399 | 1,032 |
| 1280 × 720 (HD screenshot) | 1,105 | 1,229 | 516 |
| 1920 × 1080 (full HD) | 1,105 | 1,600 | 1,548 |
| 2048 × 2048 | 765 | 1,600 | 2,322 |
| 3024 × 4032 (phone photo) | 765 | 1,600 | 6,192 |
OpenAI low-detail mode bills a flat 85 tokens at any size. Figures follow each provider's documented accounting and are estimates; actual billing can vary slightly with aspect-ratio rounding. Counting text instead? Use the text token counter or check rates on the LLM pricing table.
FAQ
Image token FAQ
How does OpenAI count image tokens?
In high-detail mode, GPT-4o-class models scale the image to fit 2048×2048, scale the short side to 768 pixels, then bill 170 tokens per 512×512 tile plus a base of 85 tokens. A 1024×1024 image costs 765 tokens. Low-detail mode bills a flat 85 tokens regardless of size.
How does Claude count image tokens?
Anthropic estimates image tokens as width × height ÷ 750 after scaling images larger than 1568 pixels on the long edge, which caps a single image around 1,600 tokens. A 1024×1024 image is roughly 1,400 tokens.
How does Gemini count image tokens?
Google bills small images (up to 384×384) a flat 258 tokens. Larger images are split into 768×768 crops, each billed 258 tokens, so a 1920×1080 screenshot costs about 1,548 tokens.
Why do image tokens matter for cost?
Image tokens bill at the same input rate as text tokens. A phone photo can cost more tokens than several pages of text, so resizing images before upload — or using low-detail mode when fine detail is unnecessary — meaningfully cuts vision API spend at volume.