Image Optimizer for AI
Resize and compress an image before you send it to Claude, ChatGPT, Gemini or Grok, and see the estimated visual-token cost before and after. Runs entirely in your browser.
Processed locally — your image is handled in your browser and is not uploaded to our servers or to any AI service.
Your image is processed entirely in this browser tab. Nothing is uploaded, and no filename, dimension or estimate is sent anywhere.
How image optimization affects AI usage
Image token usage depends on the AI provider and model. For providers with documented image-token rules, reducing image dimensions can reduce visual-token usage.
What actually changes the token count
Pixel dimensions and cropping can affect visual-token usage. Making an image smaller, or cropping away parts of it that do not matter, gives the model fewer pixels to encode. Every provider processes images differently, so the same picture can cost different amounts in different places — the estimate here always names the model tier it applies to.
How Claude counts an image
Claude currently divides an image into 28×28 pixel visual patches, one visual token per patch, with per-tier resolution limits:
- Claude High resolution (Claude 4.7 and later) — longest side up to 2,576 px, up to 4,784 visual tokens per image.
- Claude Standard (Claude models before 4.7) — longest side up to 1,568 px, up to 1,568 visual tokens per image.
Claude may automatically resize an oversized image before it counts the tokens, and this tool accounts for that when estimating savings. That is why a 4K screenshot can show a large drop in file size and no drop at all in visual tokens: the model was never going to see those extra pixels in the first place.
Why quality and format are a separate thing
JPEG and WebP compression mainly reduces the number of bytes you upload — it does not reduce visual tokens. Converting a PNG to WebP at the same width and height, or dropping the quality slider, gives you a smaller and faster upload while the token estimate stays exactly where it was.
The trade-off
Smaller images can lose fine text detail. Screenshots of code, spreadsheets, receipts and scanned documents are the usual casualties — the model cannot read what is no longer legible. Use the zoom in the comparison above to check the optimised copy before you download it, and step back up a preset if anything has gone soft.
Providers without a published rule
ChatGPT / OpenAI, Gemini, Grok do not publish a stable public formula for image tokens, so no number is shown for them. Reducing dimensions still reduces the bytes you upload, and the dimension and file-size figures above remain accurate.
These figures are estimates of visual tokens only. They do not include the text in your prompt, and they are not a price. Check your provider's own documentation and billing for authoritative numbers.
How it works
Resize an image for multimodal AI
Multimodal models read images at a limited resolution. Send something far larger and the provider quietly shrinks it before processing — you pay for the upload, wait for it, and gain nothing. Send something too small and fine text stops being readable. This tool helps you land in between, and shows the estimated cost of the choice before you commit to it.
It does not call an AI, need an API key, or upload your image. Everything happens on your device.
How image optimization affects AI usage
- Pixel dimensions and cropping can affect visual-token usage. Compression settings mostly do not.
- Providers process images differently, so there is no single universal token rule. Only providers with a documented, reproducible calculation get a number here.
- Claude currently uses 28×28 visual patches, with a maximum long edge and a maximum visual-token count that depend on the model tier.
- Claude automatically resizes oversized images, and this tool accounts for that on both sides of the comparison — so it will not claim a saving that the provider’s own resizing already took.
- JPEG and WebP quality primarily reduce upload bytes and latency, not visual tokens.
- Smaller images lose fine detail. Check small text in the side-by-side view before downloading.
Presets
- Recommended — matches the selected model’s own effective resolution, so nothing is wasted and nothing is thrown away unnecessarily.
- Low token — a more aggressive downscale when cost matters more than fine detail.
- Balanced — a middle ground for general use.
- High detail — keeps as much resolution as the model will actually use.
- Custom — your own dimensions, with aspect ratio locked by default.
Privacy
No account, no server, no uploads. The image, the optimised copy and the filename all stay on your device.
Frequently asked questions
Does making an image smaller actually reduce AI token usage?
It depends on the provider and the model. Image token usage is not universal — each provider processes images differently. For providers with documented image-token rules, reducing pixel dimensions can reduce visual-token usage. Claude publishes such a rule, so this tool can give a real number for Claude. Where a provider does not publish a stable formula, the tool says the calculation varies rather than inventing a figure.
How does Claude actually count image tokens?
Claude views an image as 28×28-pixel patches, and each patch is one visual token — so an image costs ceil(width ÷ 28) × ceil(height ÷ 28) tokens. Each model tier also caps the longest edge and the total visual tokens: the high-resolution tier (Claude 4.7 and later) allows 2576px and 4784 tokens, and the standard tier allows 1568px and 1568 tokens. This tool implements that rule directly rather than approximating it.
Why is my saving smaller than I expected?
Because Claude already downscales oversized images before charging for them, so a large original often costs no more than a medium one. A 3840×2160 screenshot and a 1920×1080 screenshot are both reduced to 1456×819 on the standard tier — identical token cost, zero saving. This tool compares what the provider would actually process in each case, not raw pixel counts, so it will tell you when there is nothing to gain instead of overstating the benefit.
Does converting PNG to WebP lower the token count?
No. If the pixel dimensions stay the same, the visual-token count stays the same. Changing format or quality reduces file size and upload time, which makes requests faster and lighter, but visual-token usage is driven by dimensions and cropping. The tool reports file-size change and token change as two separate figures for exactly this reason.
Should I crop instead of shrinking?
Often, yes. Cropping removes pixels without shrinking what is left, so text stays legible while the token cost falls. That makes it a better first move than a global downscale for screenshots and scanned documents with wide blank margins. Shrinking an image with small text can push it below the point where it can be read reliably.
Is my image uploaded anywhere?
No. Decoding, cropping, resizing and re-encoding all happen in your browser using standard canvas APIs. The image is never sent to our servers, to an AI provider, or to any analytics service, and no API key is needed. The optimised copy is created locally and saved straight from your browser.