Models · November 18, 2026 · 6 min read

What a screenshot costs your agent: image tokens on Claude, OpenAI and Gemini

Every screenshot an agent sees becomes input tokens, and each provider counts them differently. Here is the math for a 1920x1080 frame on three APIs, and why a computer-use loop pays for the same picture over and over.

Some links are affiliate links; we may earn a commission at no extra cost to you. As an Amazon Associate I earn from qualifying purchases. Disclosure

If your agent looks at things, a browser tab, a dashboard, a scanned invoice, a phone screen, then pictures are probably a bigger slice of your bill than you think. Nobody prices an image as "an image." Every provider turns it into input tokens first, and the three big APIs do that conversion in three very different ways. Same screenshot, same question, and the token count can swing by more than ten times depending on where you send it and which knob you leave at its default.

This post walks through the actual counting rules from each provider's docs, runs one ordinary 1920x1080 screenshot through all three, and then gets to the part that actually hurts: an agent loop that keeps old screenshots in its history pays for them again on every single turn.

Claude: 28-pixel patches and two resolution tiers

Anthropic's vision guide is refreshingly blunt about the math. Claude reads images in 28x28-pixel patches, each one a visual token, so an image costs ceil(width / 28) × ceil(height / 28) tokens. There's a ceiling, though, and it depends on the model. Claude 4.7 and later sit on a high-resolution tier with a 2,576-pixel long edge and a 4,784-token cap. Older models sit on a standard tier capped at 1,568 pixels and 1,568 tokens. Anything bigger gets scaled down to fit.

So a 1920x1080 screenshot costs 1,560 tokens on a standard-tier model (it gets shrunk to 1456x819) but 2,691 tokens on a high-resolution model, because it fits as-is. Anthropic turns the high-resolution tier on automatically for those models, no header or opt-in, and the docs pitch it for computer use, screenshot understanding and dense documents. It also means the same screenshot got about 70% more expensive when you upgraded.

The guide does the dollar math itself: at Claude Opus 5.5's $4 per million input tokens, a 1000x1000 image runs about $5.18 per thousand images and a 4K image about $19.14 per thousand. Our 1920x1080 frame lands around $10.76 per thousand on Opus 5.5. On Claude Haiku 5.5 at $0.10 per million, the same frame is about 27 cents per thousand, which is a good reminder that model choice moves the number far more than image size does.

Two limits matter for agents in particular. A request can carry up to 600 images (100 on models with a 200k context window), but once a request holds more than 20 images, a stricter per-image dimension limit kicks in, and the docs say images from earlier turns and screenshots nested inside tool_result blocks count toward that 20. Their safe advice is to keep each side under 3,000 pixels. And for computer-use and browser-use screenshots returned as tool results, the API doesn't quietly shrink an oversized image. It rejects it. Resize on your side before you hand it back.

OpenAI: 32-pixel patches, a detail knob, and a multiplier

OpenAI's images and vision guide uses a similar idea with different numbers. Newer models cover the image with 32x32-pixel patches, apply a per-model resizing budget, then multiply the patch count by a model-specific multiplier to get billable tokens. For gpt-6-astra that multiplier is 1.2.

The knob that matters is detail. It accepts low, high, original or auto, and if you leave it out it defaults to auto. On gpt-6-astra, low fits the image inside 512x512 pixels, high allows up to 2,500 patches, and auto behaves like original, which keeps the image's real dimensions (only things wider than 65,535 pixels get scaled) and rejects anything over 30,000 patches outright instead of resizing it.

Run the 1920x1080 frame through high and you get 60 × 34 = 2,040 patches, under the budget, so no resize, times 1.2 gives 2,448 tokens. At GPT-6 Astra's $10 per million input, that's about $24.48 per thousand screenshots. Switch to low and the frame shrinks to 512x288, which is 16 × 9 = 144 patches and 173 billable tokens, roughly $1.73 per thousand. That's a fourteen-fold difference from one parameter.

Don't flip everything to low just because it's cheap. The guide itself says to use original for OCR, small objects and computer use, because those need the fine detail and accurate coordinates. It also notes that on some older models like gpt-5.4, low actually has a bigger budget than high and can cost more. Check the sizing table for your exact model before you assume.

Gemini: a fixed budget per image

Google went a different way. On Gemini 4 and Gemini 3 models, the media resolution page gives each image a fixed token allocation set by media_resolution: 280 tokens at low, 560 at medium, 1,120 at high and at the default, and 2,240 at ultra_high. You can set it per image, so one request can send a dense diagram at high and a throwaway context photo at low. Google specifically recommends ultra_high for computer use.

Because the budget doesn't change with pixel count, the 1920x1080 frame costs 1,120 tokens at the default. On Gemini 4 Carbon at $2 per million input, that's about $2.24 per thousand, or $4.48 at ultra_high. The image understanding guide adds the practical limits: inline image data caps the whole request at 20MB, past that you upload through the Files API, and a request can hold up to 3,600 image files.

The real cost is the loop, not the picture

Single-image prices look tiny. The trouble starts when an agent takes a screenshot, acts, takes another screenshot, and keeps all of them in the conversation history. Every model call resends the full history, so screenshot number one gets billed on turn one, turn two, turn three and on until the run ends.

Here's what that does. Say a computer-use run takes 30 steps, each adds one 1920x1080 screenshot, and you're on Claude Opus 5.5 at 2,691 tokens per frame. Keeping every screenshot means the image tokens sent across the run add up to 1 + 2 + ... + 30 = 465 screenshot-sends, or about 1.25 million input tokens. That's roughly $5.01 per run in images alone, before any text or output. Keep only the last three screenshots and the total drops to 87 screenshot-sends, about 234,000 tokens or around $0.94. Same task, same model, about a fifth of the cost, and the model usually doesn't need a picture of a page it left ten steps ago anyway.

Trimming old images also keeps you away from Claude's more-than-20-images rule, which counts history and tool results. A run that never prunes crosses that line at step 21, and from then on every image in the request has to meet the stricter size limit whether you planned for it or not.

If you'd rather keep history intact, prompt caching cuts the price of the unchanged prefix, and context compaction can swap old screenshots for a one-line text note of what was on screen. The most effective version is usually the dumbest: replace each old screenshot with a sentence like "Checkout page loaded, total showed $42.10" and drop the image.

What to actually do

Resize before you send. A 1280x800 frame on Claude is 46 × 29 = 1,334 tokens, about half the 1920x1080 cost, and for most UI work it's plenty readable. If you do coordinate work, resize to a size you control and map clicks back yourself, because both Anthropic and OpenAI return coordinates relative to the image the model actually processed, not your original.

Set the detail or resolution knob on purpose. OpenAI's default is auto, which on newer models means full original size. Gemini's default is 1,120 tokens. Neither default is wrong, but neither is tuned for your task. Use the cheap setting for "did the page load" checks and save the expensive one for reading fine print.

Crop when you can. If the agent only needs the error toast in the corner, send the corner. And measure. Log the image token count your provider reports per call for a day of real traffic, and you'll know whether screenshots are a rounding error or your biggest line item.

If you want a deeper grounding in how multimodal inputs fit into production LLM systems, AI Engineering by Chip Huyen covers the bigger picture well. For current per-token prices on every model mentioned here, see our models page.

Newsletter

The daily agent toolkit.

New tools, MCP servers and frameworks, in one short email every morning. Free, no spam. Sign up and get the free AI agent cost cheat sheet (PDF).