Image Token Calculator
A 1920 × 1080 image costs 1,560 input tokens on Claude Haiku 4.5 (resized to 1456 × 819, ⌈w/28⌉ × ⌈h/28⌉), 2,691 on Claude Opus 5.5, 1,105 on GPT-4.1 at high detail and 1,120 on Gemini 3, per the vendor docs checked 16 October 2026.
How to use
- Pick a common size or type the image width and height in pixels.
- Enter how many images you will send and choose the detail level (high or low).
- Read the table: tokens per image, total tokens and input cost for each model. A size in brackets is the size the model resizes the image to.
How it works
OpenAI patch models (GPT-6 Astra, GPT-5.5): the image is fitted into the detail box (512 × 512 for low), then covered with 32 × 32 pixel patches. If the count is over the patch budget (2,500 at high), the image is shrunk until it fits. Tokens = patches × 1.2, rounded up. Example from the OpenAI guide: 2048 × 2048 at high becomes 1600 × 1600 = 2,500 patches = 3,000 tokens.
OpenAI tile models (GPT-4.1, GPT-4o mini): at low detail an image costs only the base tokens. At high, it is fitted into 2048 × 2048, the shortest side is scaled to 768, and the 512 pixel tiles are counted. Tokens = base + tiles × tile tokens (85 + 170 per tile for GPT-4.1, 2,833 + 5,667 for GPT-4o mini). 1024 × 1024 becomes 768 × 768, 4 tiles, 765 tokens.
Claude: tokens = ⌈width ÷ 28⌉ × ⌈height ÷ 28⌉. Before counting, the image is shrunk to the largest size that keeps both sides at or under 1568 px and the cost at or under 1,568 tokens (Haiku 4.5), or 2576 px and 4,784 tokens on Claude 4.7 and later. The tool runs the reference implementation from the Anthropic docs, so 1920 × 1080 is 1456 × 819 = 1,560 tokens on the standard tier and 2,691 tokens on the high resolution tier.
Gemini 3: a fixed count per media_resolution level, 280 tokens at low and 1,120 at the default and high levels. Gemini 2.5: an image with both sides at or under 384 px is 258 tokens, a larger one is split into 768 × 768 tiles of 258 tokens each.
Cost = tokens × number of images × the standard input price per million tokens from the LLM API cost calculator price table. Sources: developers.openai.com/api/docs/guides/images-vision, platform.claude.com/docs/en/build-with-claude/vision and vision-coordinates, ai.google.dev/gemini-api/docs/media-resolution and docs/tokens, all checked 16 October 2026.
Tokens for common sizes
| Size | GPT-5.5 high | GPT-4.1 high | Claude Opus 5.5 | Claude Haiku 4.5 | Gemini 3 |
|---|---|---|---|---|---|
| 1024 × 1024 | 1,229 | 765 | 1,369 | 1,369 | 1,120 |
| 1920 × 1080 | 2,448 | 1,105 | 2,691 | 1,560 | 1,120 |
| 3840 × 2160 | 2,765 | 1,105 | 4,784 | 1,560 | 1,120 |
How to send images for less
- Resize before you upload. Every model above caps the size it actually reads, so pixels past that cap cost upload time and nothing else.
- Use low detail on OpenAI models when you only need the gist of a photo. At low a GPT-4.1 image costs 85 tokens whatever its size.
- Crop to the part that matters. A screenshot of one dialog costs far less than the full 4K screen, and the model reads small text better.
FAQ
How many tokens is a 1024 × 1024 image?
About 1,229 on GPT-5.5 and GPT-6 Astra at high detail, 765 on GPT-4.1, 1,369 on Claude and 1,120 on Gemini 3 at the default resolution.
Is the old Claude rule of width × height ÷ 750 still right?
No. The current Anthropic docs count 28 × 28 pixel patches. For a 1000 × 1000 image that gives 1,296 tokens, where the old rule gave 1,333.
Why does a 4K image cost the same as 1080p on some models?
Because the model shrinks it first. Claude Haiku 4.5 reduces both 1920 × 1080 and 3840 × 2160 to 1456 × 819, and GPT-4.1 scales both to 1365 × 768 before it counts tiles.
Is the cost exact?
The token count follows the published rules. The cost uses the standard input price and leaves out batch discounts, cached input and the text in your prompt. Floating point rounding on the provider side can move a count by one token.