Example 1A Food Blogger's Recipe Card (1024x1024 on GPT-4o High)
Given:1024 × 1024, GPT-4o, high
Rezultat:~765 tokens, costing about $0.0038 per image
Let's say you are uploading a square photo of a handwritten recipe card. The AI splits this 1024x1024 image into four 512x512 tiles. It charges a base fee of 85 tokens just to look at the image, plus 170 tokens for each of the 4 tiles. That's 85 + (4 × 170) = 765 tokens. At a rate of $5 per million tokens, you're paying less than half a cent per image!
Example 2Quick Receipt Scanning (1024x1024 on GPT-4o Low)
Given:1024 × 1024, GPT-4o, low
Rezultat:85 tokens flat, costing about $0.0004 per image
If you just need to classify the receipt type (like 'grocery' or 'gas') without reading the tiny font, you can switch to Low Detail mode. The AI bypasses the tile system entirely and charges a flat rate of 85 tokens, regardless of the physical size. This is nearly 10 times cheaper than high-detail mode!
Example 3A Student's Whiteboard Notes on Claude 3 (2048x2048)
Given:2048 × 2048, Claude 3
Rezultat:~5,592 tokens, costing about $0.0168 per image (at Sonnet rates)
Claude 3 takes a different approach. Instead of chopping the image into tiles, it estimates tokens using a simple ratio of pixels divided by 750. For a massive 2048x2048 canvas, the math is (2048 × 2048) / 750, which equals roughly 5,592 tokens. Since Claude charges more tokens for larger dimensions, resizing this whiteboard photo before uploading can save you a bundle!
Example 4Trimming a Heavy 4K Smartphone Photo (3840x2160 to 1024x576)
Given:Before: 3840×2160 GPT-4o High (~2,720 tokens) | After: 1024×576 GPT-4o High (~595 tokens)
Rezultat:Over 4.5× cost savings just by shrinking the image dimensions!
A raw 4K screenshot requires 32 tiles! By simply resizing the image down to a standard 1024-pixel width before sending it to the AI, you drop from 2,720 tokens down to just 595 tokens. The AI can still easily read the main text and shapes, but you save 78% on your bill.