Tool
Image Token Calculator
The same image can cost 1,120 tokens on one vendor and over 14,000 on another — because each vendor counts images differently. Enter the pixel dimensions and see all three side by side, before you send anything.
What it costs
Enter dimensions to compare.
| Model | Tokens / image | Resized to | $ / image | $ / 1,000 images |
|---|
Costs use this site's model prices (see the context window table). Counts are estimates from the vendors' documented formulas, not billed counts: OpenAI notes billing can differ by one token, and Claude's exact downscale rule is approximated here. For an exact count, use OpenAI's calculator, Anthropic's count endpoint, or Gemini's countTokens API.
Why one image has three prices
Text tokens are (roughly) shared across vendors. Image tokens are not: each vendor runs the image through its own encoder first, and the encoder — not the price list — decides the count. A phone photo is 14,364 tokens on GPT at auto detail, 4,740 on Claude, and 1,120 on Gemini at high resolution. At the same $2 per million input tokens that would be $28.73, $9.48 and $2.24 per thousand images. The lever is the formula, not the rate.
How each vendor counts
OpenAI (GPT-6 family). The image is covered in 32×32-pixel patches and each patch costs 1.2 tokens, rounded up. At low the image is first shrunk to fit 512×512. At high it is shrunk to fit a 2,500-patch budget (and a 65,535-pixel max dimension). At auto — the default — it keeps its size, except images over 65,535 pixels on a side are scaled down; if the result still needs more than 30,000 patches, the API rejects the request instead of resizing it.
Claude (5.5 family). The image is cut into 28×28-pixel patches, one token each. High-resolution models accept a 2,576-pixel long edge and up to 4,784 tokens; anything larger is downscaled, preserving aspect ratio, until it fits both limits. A 1920×1080 screenshot needs no resizing: 2,691 tokens. A 4K frame is downscaled to 2576×1449 and stops at exactly 4,784.
Gemini (3 family). No patch math at all: each image costs a flat budget set by media_resolution — 280 at low, 560 at medium, 1,120 at high, 2,240 at ultra_high. PDF pages use the same budgets per page. That makes Gemini the only vendor where resizing the image saves nothing; the knob is the resolution setting, not the dimensions.
Cutting the bill
On OpenAI and Claude, shrink before you send: a phone photo resized to 1024×768 drops from 14,364 to about 922 tokens on GPT at auto detail, because the patch count falls with the area. On Gemini, don't bother resizing — drop from high to medium resolution instead and halve the count. And if an image is sent on every turn of a long chat, it is re-billed on every turn: that per-image number is multiplied by the number of turns, not paid once.
Next: paste the prompt that goes with the image into the Token Counter for the full request cost, or check the context window table for the model prices used above.