Multimodal Token Calculator
How many tokens does an image, audio clip, or PDF cost? OpenAI, Anthropic, and Google count multimodal tokens differently — sometimes by 10× for the same image.
| Provider | Image tokens | Audio tokens | PDF tokens | Total tokens | Notes |
|---|---|---|---|---|---|
| OpenAI | 2,125 | — | — | 2,125 | high detail |
| Anthropic | 2,765 | — | — | 2,765 | w×h/750 |
| 258 | — | — | 258 | fixed tiling |
Token counts are approximations based on official provider documentation. Actual counts may vary by model version. Verified 2026-08-20.
How to Use
Enter image dimensions
Enter width and height in pixels. Choose high or low detail for OpenAI models. The token count varies dramatically by provider for the same image.
Set media counts
Number of images, audio length in seconds, and PDF pages. Tokens are calculated for each media type separately.
Compare providers
The table shows token count per provider side-by-side. Google Gemini is typically cheapest for images. OpenAI high-detail is most expensive.
Calculate cost
Multiply token counts by your model's per-million-token input price to get cost per API call.
Why multimodal token counts vary so much
Each provider uses a different algorithm. OpenAI tiles images and charges per tile. Anthropic uses (width × height) / 750. Gemini uses a fixed 258 tokens for most images. The same 1920×1080 image costs 2,125 tokens on OpenAI, 2,765 on Anthropic, but only 258 on Gemini — a 10× difference. This makes Gemini dramatically cheaper for vision workloads at scale.
How to Cut This Cost
Resize images before sending. OpenAI GPT-4 Vision charges per 512×512 tile — a 4K image costs 16× more than a 1K image with often no quality benefit for most tasks.
Use low-detail mode where layout is irrelevant. OpenAI low-detail fixed cost is 85 tokens vs up to 1,500+ for high-detail.
Convert images to text (OCR) first for document tasks. Text is far cheaper than image tokens for content that is purely textual.
FAQ
Have more questions? Contact us
Last verified: 2026-08-20 · methodology · data sources