Claude bills vision by pixel area, so compressing the image saves nothing

Anthropic's vision documentation prices an image as a function of its pixel dimensions: tokens are approximately width in pixels times height in pixels, divided by 750 ^[1]. File size does not appear in that expression, and neither does format ^[2], which means the first lever most teams reach for when an inference bill grows is wired to nothing.

A 4 MB PNG and a 300 KB JPEG of the same 1200x800 scene cost the same number of tokens ^[3]. Compressing harder saves upload bandwidth and saves nothing on the bill; resizing saves both ^[3]. The divisor is round because it approximates a patching step, and the total scales with area rather than with either edge ^[4]. The consequence worth internalising: doubling both edges quadruples the token count, so a "slightly bigger" screenshot is not slightly more expensive ^[4].

The worked numbers, as set out in a dev.to walkthrough of Anthropic's documentation. A 640x480 UI screenshot is 307,200 pixels, or about 410 tokens ^[5]. A 1200x800 web photo is 960,000 pixels, about 1,280 tokens ^[6]. A 4032x3024 phone photo is 12,192,768 pixels, which the raw formula puts at 16,257 tokens ^[7] and which you are not charged, because two documented limits intervene. If the long edge exceeds 1568 pixels, the image is scaled down before it reaches the model ^[8]. And an image is scaled so that it does not exceed roughly 1,600 tokens, which working the formula backwards is about 1.15 megapixels ^[9], derived from 1,600 times 750 ^[10].

Applied to the phone photo, the first limit gives 1568x1176, which is still 2,458 tokens and above the ceiling ^[11]; a second reduction to about 1238x928 lands at roughly 1,530 tokens ^[12]. The exact output dimensions are Anthropic's choice, not yours, and those intermediate numbers are an illustration of the documented ceiling rather than a published table ^[13]. Net effect: the 12-megapixel upload ends up costing about a tenth of what its pixel count implies ^[14], and a 12-megapixel and a 2-megapixel photo of the same scene cost roughly the same ^[15]. What you lose is detail, discarded by a downscale that has no idea what you needed to see ^[16]. If the model has to read small text, crop to the region rather than sending the whole frame ^[17].

This reshapes document pipelines more than it looks. An A4 page scanned at 300 dpi is about 2480x3508 pixels ^[18]; sent whole, it costs the ceiling of about 1,600 tokens ^[19], which is roughly one seventh of what its area would otherwise imply ^[20], and the quadrant you actually care about arrives at a fraction of the resolution you scanned ^[19]. Below the ceiling the meter is honestly proportional: 640x480 really is a third of the cost of 1200x800, and there is no minimum charge that makes small crops pointless ^[21].

Two things to watch. Anthropic calls the formula an estimate; the billed figure is what comes back in usage.input_tokens, and the pre-flight figure is the count_tokens endpoint, which accepts image blocks ^[22]. And image handling is an area vendors revise, so re-check the limits before you build a budget on them ^[23].