Image Vision Token Calculator
Drop an image here
JPG, PNG, or WebP — drag & drop, click, or paste with Ctrl+V
100% private — stays in your browser, never uploaded
Up to 15MB
Calculate
Sample result — a 2048×1536 photo. Drop your own image, or try an example, to replace it.
GPT-4o · High detail
2,125tokens
12 tiles × 170 + 85 base · $0.005313 per image · $5.31 for 1,000
Use Low for classification, High for reading fine text or OCR.
| Model | Tokens | Cost/image | Cost/1,000 |
|---|---|---|---|
| GPT-4o · High | 2,125 | $0.005313 | $5.31 |
| GPT-4o · Low | 85 | $0.000213 | $0.2125 |
| GPT-4o mini · High | 2,125 | $0.000319 | $0.3187 |
| GPT-4o mini · Low | 85 | $0.000013 | $0.0127 |
| Claude Sonnet 5 | 4,194 | $0.008388 | $8.39 |
| Gemini 3.7 Flash | 1,548 | $0.001161 | $1.16 |
OpenAI's tiling is documented and exact; Claude and Gemini figures use each provider's own published approximation — see the FAQ below. Approximate — verified against each provider's official pricing page as of August 31, 2026. Providers change prices without notice; check their page before a billing decision.
Dimensions are read locally via the browser's Image() object — the image itself is never uploaded.
What is a vision token? How do images cost tokens?
Vision-capable models charge for an image the same way they charge for text — in tokens — but the conversion has nothing to do with characters. Unlike text, where roughly four characters make one token, an image's token cost is a function of its pixel dimensions and the requested detail level. GPT-4o and GPT-4o-mini use a documented tile system: the image is divided into 512×512 squares, and each square costs a fixed number of tokens. Claude and Gemini don't tile the same way, but both scale roughly with pixel count too — see the per-model formulas below.
How many tokens does an image cost? (by model and size)
| Model | Low detail | 1024×1024 high | 2048×1536 high |
|---|---|---|---|
| GPT-4o | 85 tokens fixed | 765 tokens (4 tiles) | 2,125 tokens (12 tiles) |
| GPT-4o-mini | 85 tokens fixed | 765 tokens (4 tiles) | 2,125 tokens (12 tiles) |
| Claude Sonnet | ≈ (width × height) / 750 — about 1,398 tokens at 1024×1024, 4,194 at 2048×1536 | ||
| Gemini Flash | 258 tokens flat up to 384×384, else 258 × ceil(w/768) × ceil(h/768) tiles | ||
GPT-4o pricing: $2.50 / 1M input tokens. GPT-4o-mini: $0.15 / 1M. Approximate — verified against each provider's official pricing page as of August 31, 2026. Providers change prices without notice; check their page before a billing decision.
The GPT-4o high detail formula, step by step
- 1. Scale the image to fit within 2048×2048, preserving aspect ratio.
- 2. Scale the shortest side to 768px.
- 3. Count 512×512 tiles:
tiles = ceil(width / 512) × ceil(height / 512). - 4.
tokens = tiles × 170 + 85.
Example: a 2048×1536 image already fits step 1, scales to roughly 1024×768 in step 2, tiles to 2×2 = 4 in step 3, and costs 4×170+85 = 765 tokens. This calculator's grid and headline figure use the tile count directly on the raw dimensions — the textbook version of the formula everyone searches for — rather than simulating the pre-tiling resize step, so very large images may bill slightly lower in a live API call than the number shown here.
Low vs high detail — when to use which
- Low detail — 85 tokens fixed, the image resized to fit inside 512×512 internally. Good for "is there a cat in this photo?" style classification. Up to 25× cheaper than high detail.
- High detail — tiles × 170 + 85, tiled at full resolution. Needed for OCR, reading small text, or any answer that depends on fine detail.
Rule of thumb: use low detail if the model just needs to recognize what's in frame; use high if it needs to read something in it.
How to reduce image token cost by up to 80%
- Resize to 768px on the short side. Cuts a 12-tile image to 4 tiles — about a 64% saving, still high detail quality.
- Use low detail when possible. 85 vs 2,125 tokens is a 96% saving for pure classification tasks.
- Crop to the relevant area. Don't send a full screenshot when only a corner of it matters.
- File compression doesn't help. Quality/WebP compression shrinks the file, not the pixel dimensions — tiling only looks at pixel dimensions, so it saves zero tokens.
- Batch-resize before the API call. Resize every image to 768px up front rather than per-request.
- Use GPT-4o-mini for vision when it can do the job. Identical token count to GPT-4o, at roughly 1/16th the input price.
For a hands-on resize/crop/OCR workflow on your own image rather than just the numbers, see the Image to LLM Optimizer.
How to estimate cost for 1,000 images
Total cost = tokens per image × price per 1M tokens / 1,000,000 × number of images. Example: 1,000 images at 1024×1024, high detail, on GPT-4o — 765 tokens × $2.50 / 1,000,000 × 1,000 ≈ $1.91. The Batch tab above runs this for any image, model, and count.
Related tools
- Image to LLM Optimizer — once you know the number, this is where you fix it: resize, crop, or OCR the same image down live, with before/after savings.
- Token Counter — count tokens and cost for text you already have, across nine models.
- File to Tokens — a PDF with scanned pages carries vision tokens too; this estimates the whole document.
- Screenshot to LLM — need code generated from a UI screenshot, not just its token cost?
Frequently asked questions
How many tokens does an image cost in GPT-4o?
At low detail it is a flat 85 tokens, whatever the resolution. At high detail it is tiles × 170 + 85, where tiles = ceil(width / 512) × ceil(height / 512). A 1024×1024 image is 2×2 = 4 tiles = 765 tokens. A 2048×1536 image is 4×3 = 12 tiles = 2,125 tokens. Drop your own image into the calculator above to see the exact figure for its real dimensions.
What is the difference between low and high detail?
Low detail resizes the image to fit inside 512×512 internally before the model ever tiles it, so it always costs 85 tokens — good for classification tasks like "is there a person in this photo?" High detail tiles the image at its real resolution so the model can read fine print, at 170 tokens per 512px tile plus an 85-token base. High detail typically costs 10–25× more than low, so the choice matters for both accuracy and cost.
How does GPT-4o count 512px tiles?
OpenAI's documented pipeline first scales the image to fit within 2048×2048, then scales its shortest side to 768px, then divides what's left into 512×512 tiles — each tile costs 170 tokens, plus an 85-token base. This calculator's tile grid overlay draws that division directly on your image so you can see it rather than just read the formula.
How many tokens for Claude and Gemini vision?
Neither publishes an OpenAI-style tile formula. Anthropic's own docs give an approximation of tokens ≈ (width × height) / 750 for Claude. Google documents that a Gemini image up to 384×384 is a flat 258 tokens, and anything larger is tiled into up to 768×768 crops at 258 tokens each. Both are shown alongside GPT-4o and GPT-4o-mini in the model table above for the same image.
How do I reduce vision token cost?
Resize before you send it — cutting a 2048px-wide image to 768px on its short side typically drops 12 tiles to 4, about a 64% saving, while staying in high detail for readable text. Use low detail (85 tokens flat) for anything that only needs classification, not reading. Cropping to the part of the frame that actually matters shrinks the tile count further. For a full resize/crop/OCR workflow rather than just the numbers, see the Image to LLM Optimizer.
How much do 1,000 images cost?
Multiply one image's token count by the model's per-token price and by the image count. A 1024×1024 image at GPT-4o high detail is 765 tokens; at $2.50 per 1M input tokens, 1,000 of them cost 765,000 tokens × $2.50 / 1,000,000 ≈ $1.91. The batch calculator above does this for any image count and any of the four models.
Is my image uploaded anywhere?
No. The image is read locally with the browser's own Image() object to get its pixel dimensions — nothing is sent to Toolsda's servers or any API. Every number on this page, including the tile grid, is computed from those dimensions in your browser tab.