Your team budgets vision spend for a product photo feature on the Claude API. Claude estimates image tokens as (width in px times height in px) divided by 750, and only downscales images whose long edge exceeds 1568 px. A user uploads a 1092x1092 photo. Predict the approximate token count Claude charges for this image (round to the nearest 10).
Long edge is 1092 px which is below the 1568 px downscale threshold, so the formula applies directly: 1092 times 1092 divided by 750 equals about 1590 tokens.
Imagine a calculator with a simple rule: count the dots in your picture (width times height), then split that count into bundles of 750. Each bundle is one token. The calculator only shrinks the picture first if its longest side is bigger than 1568 dots wide. This picture is 1092 by 1092, so the calculator skips shrinking. Multiply: 1092 times 1092 gives about 1.19 million dots. Divide by 750: about 1590 bundles, which is about 1590 tokens. Knowing this number lets a team set a per-image budget instead of being surprised by the bill at the end of the month.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Image token cost looks like trivia until your monthly invoice arrives 4x over estimate. The Claude formula is the cleanest of the major providers because it is linear in pixel area, but using it correctly requires checking the downscale condition and budgeting at the feature level, not the request level.
This deep dive walks through the calculation, the cost levers that follow from it, and the comparison to other providers.
The formula and the downscale rule
Claude estimates input image tokens as:
With a single conditional: if the image's long edge exceeds 1568 px, Claude downscales the image so the long edge equals 1568 px (preserving aspect ratio), then applies the formula. The downscaling happens server side; you cannot send a higher resolution and somehow get the model to see more pixels than the cap allows.
For a 1092x1092 image:
- Long edge is 1092 px, below 1568. No downscale.
- Pixel area is 1092 times 1092 = 1,192,464 px squared.
- Tokens are 1,192,464 / 750 = 1589.95.
- Rounded to the nearest 10: 1590 tokens.
This is near the per-image cap before downscaling: a maximum-allowed 1568x1568 square image is 1568 times 1568 / 750 = 3,278 tokens. So 1092x1092 sits in the upper-mid portion of the cost band.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Anthropic's vision pricing page documents the width times height divided by 750 formula and the 1568 px downscale point for Claude Opus 4.7 and Sonnet 4.6
- OpenAI's GPT-5.5 vision pricing uses a tile-based formula (base tile plus detail tiles), not a single area formula, so cross-provider comparison requires unit conversion
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you estimate monthly vision spend for a product launch?
Multiply expected requests per day by average images per request, by tokens per image (from this formula), by input price per 1M tokens, by 30 days. Add a buffer for the launch spike. Track per-feature image budgets so a single team cannot blow the line item without it showing up.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Multiplying width and height in centimeters or guessing a round 1500 without doing the division by 750; or applying the downscale rule when the image is already below the threshold.
60 second bullets to scan on the way to the call.
Claude's image token formula and the downscale threshold
Why client-side downscaling is the primary cost lever
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.