← Back to articles

Image understanding · API billing

Estimate the input cost of sending an image to a vision model

A model's USD-per-million-input-tokens rate does not reveal the price of an image by itself. Each provider converts pixels or a media-resolution choice into tokens differently.

Published Nov 9, 2026Official sources checked Nov 9, 2026

Illustrative standard-rate estimates

One 512 × 512 image, three different token rules

These are image-input-only examples, not comparable image-quality benchmarks or complete request prices. The models have different processing rules and capabilities.

Input rates above are standard non-cached list rates for the named routes, checked Nov 9, 2026. They exclude text prompts, output tokens, system instructions, cache differences, regional premiums, retries, and additional processing. Values are estimates, not invoice guarantees.

Correct comparison method

Why the same image has different token counts

  1. OpenAI: Some models use 32px patches, per-model multipliers, and detail-specific resizing. Others use base-plus-tile accounting. Use the official image input calculator for an exact model and setting.
  2. Anthropic: 28px patches apply after any necessary downscaling. Larger images can exceed model-specific native resolution or visual-token limits; the simple raw-dimensions formula then overestimates.
  3. Google Gemini: Gemini 3 image token accounting depends on media_resolution. Low / medium / high / ultra-high correspond to different token budgets; previous Gemini generations behave differently.
  4. Complete request: Image-input tokens × input rate + text-input tokens × input rate + generated tokens × output rate. Check the exact provider billing response whenever possible.

Image understanding is not image generation and is not equivalent to page-priced OCR. The site's token calculator only estimates correctly after you supply an appropriate total input token count; it does not compute model-specific image tokens automatically.