TokenSave Data · Dataset

How many tokens does an image cost? Claude vs Gemini vs GPT

A 12-megapixel phone photo costs 1,065 input tokens on Gemini, 4,005 on Claude 5.5 and 14,364 on GPT-6 Astra with default settings. Gemini charges about the same for every image, so it is the cheapest for large pictures and the most expensive for small ones.

Measured 2026-10-18 · 10 resolutions · 6 model settings · License CC BY 4.0 · Related: token cost by language, by data format

DOI 10.5281/zenodo.23443906 · Also on Hugging Face · Kaggle · GitHub

Key findings

• Gemini charges a flat rate. Every image we sent cost 1,065 to 1,101 tokens on Gemini 3.x, from a 256×256 icon to a 12 MP photo. The small icon costs 1,090 tokens there against 77 on GPT-6 Astra, but the photo costs 93% less than on GPT-6 Astra.
• GPT-6 Astra does not shrink images by default. Its default detail level keeps the original size, so cost grows with every pixel: 14,364 tokens for a 12 MP photo. Setting detail: "high" caps it at 2,942 tokens (-80%).
• Claude 5.5 keeps more detail than Haiku 4.5, and charges for it. Up to about 1 megapixel the two cost almost the same (1,446 vs 1,434 tokens at 1024×1024). Above that Haiku 4.5 levels off near 1,661 tokens, while Claude 5.5 needs 70% more for a 1080p screenshot and 141% more for a 12 MP photo.
• Resizing before you send saves money on Claude and GPT. A 1024×1024 image costs 1,446 tokens on Claude 5.5 and 1,229 on GPT-6 Astra; shrinking a 12 MP photo to that size cuts its cost by 64% on Claude 5.5 and 91% on GPT-6 Astra. On Gemini resizing changes almost nothing.

All resolutions

Input tokens for one image. Claude and Gemini are measured; GPT columns are computed with OpenAI's published formula.

Method

• One synthetic PNG (colour gradient with shapes) rendered at 10 sizes, from 256×256 to 4032×3024. A second image with different content was measured at three sizes on Claude 5.5 and Gemini and gave identical counts, so the counts depend on size, not content.
• Claude: each image sent through Claude Code (claude -p, stream-json input) with a fixed one-line system prompt and a one-character text message; reported input tokens minus the same run without the image. Opus 5.5 and Sonnet 5.5 gave identical counts. The figures include the small per-image overhead of a message.
• Gemini: per-image token count shown by Google AI Studio for Gemini 3.8 Flash at the default media resolution.
• GPT: computed with the 32×32-pixel patch formula and model limits in OpenAI's vision guide (patch count × 1.2). Our implementation reproduces the guide's three worked examples exactly. Not measured through the API.

Limitations

• Input tokens only. The price per token differs by model, so equal token counts do not mean equal cost.
• Default settings. Gemini's media resolution and GPT's detail option change the counts; we show GPT-6 Astra both ways.
• Claude counts are from Claude Code; the API may differ by a few tokens of message overhead.

How to cite

Free to use, quote and republish with attribution under CC BY 4.0. Please link to this page.

Embed the chart

Related

• Token cost by language on Claude, Gemini and GPT
• Token cost by data format (JSON vs CSV vs TOON)
• Count tokens for your own text

Questions or corrections: contact@jonhisking.com