why images cost way more than text in ai (and when it matters).

elisabeth hitz · july 14, 2026 · 4 min read

a picture of a page can cost an ai far more than the exact same words handed over as plain text, because the model has to process the image as a grid of visual tokens rather than a tidy string of characters. you are paying more and getting a less reliable read at the same time. once you feel why that is true, the habit fixes itself: prefer text, and save images for when the visual itself is the point.

to see the intuition, you have to know what a token is. a model does not read letter by letter. it chops your input into chunks called tokens, and you are billed by how many it has to handle. text is extremely efficient here. a paragraph becomes a short, clean run of tokens.

why does an image cost so much more?

an image does not get chopped into words, because it has no words to chop. it gets carved into a grid of little tiles, and each tile becomes its own visual token. a full page rendered as an image can turn into a large block of these tiles, so the same content you could have pasted as a compact strip of text arrives instead as a heavy grid. that is why a page-image can cost noticeably more than its words as text, sometimes many times more depending on the size and detail of the image.

i am not going to hand you a precise multiplier, because it swings with resolution and the model. the honest version is the one that matters: for a page that is mostly words, the image form is the expensive form, and it is expensive for no benefit, since the words were the whole point.

it is not just cost, it is reliability

the second half of the trade is accuracy. reading text off a grid of pixels is a harder, fuzzier job than reading real characters, so it is more error-prone, and the errors hide in exactly the places you care about. numbers, tables, ids. i went deep on that misread problem in stop screenshotting text for ai. so the image form is both pricier and shakier. text wins on both counts when text is what you have.

when does the cost actually matter?

be honest about scale, because the answer is not always. here is where it bites and where it does not.

• it matters on long documents. a fifty-page scan as images can balloon your token count and your bill. the same pages as text stay lean.
• it matters on batch work. if you are running the same job over a hundred files, a per-file overcharge multiplies fast into real money and slower runs.
• it matters on tables and numbers. here the reliability hit stacks on top of the cost hit, so you lose twice.
• it barely matters for a one-off quick look. pasting a single screenshot to ask one question is fine. the overhead is trivial at that size and not worth fussing over.

the takeaway

the rule is simple and it holds everywhere. prefer text. it is cheaper, it is more accurate, and it is easier for the model to use. reserve images for when the visual itself is the information, a chart to read, a design to react to, a photo to describe, a bug on screen. in those cases the pixels carry meaning that no text could replace, and the cost is worth it.

the rest of the time, get your content into text and keep it clean. if it is trapped in a scan, run ocr first, then format the result the way the pillar on markdown vs pdf for claude lays out. cheaper input and better answers are usually the same move.

faq

how much more does an image cost than text?

it depends on the image size and the model, so there is no single number. the safe intuition is that a page-image can cost several times what the same words cost as plain text, and sometimes much more.

if i only send one screenshot, does it matter?

not really. for a single quick question the overhead is small and not worth worrying about. the cost adds up on long documents, batch jobs, and anything you repeat at scale.

so should i never send images?

no. send images when the visual itself is the point, like a chart, a design, or a ui bug. just do not send a picture of plain text when you could send the text.