Vision Token Cost in 2026: Image Resolution Is the Price

Q: How many tokens does an image use in Claude?

Anthropic's vision documentation says Claude splits an image into 28 by 28 pixel patches and charges one visual token per patch, so the cost is the ceiling of width over 28 times the ceiling of height over 28. Images above the model's cap are downscaled first: 1,568 px and 1,568 tokens on the standard tier, 2,576 px and 4,784 tokens on the high-resolution tier.

Q: Is the width times height divided by 750 formula still right for Claude?

No. That rule of thumb was widely quoted, but it is not what Anthropic's current vision documentation describes. The current rule counts 28 by 28 pixel patches and caps the total per resolution tier. Use the documented patch rule, or the token counting endpoint, and re-check the rule whenever you move to a new model version.

Q: How does OpenAI calculate image tokens in 2026?

OpenAI documents two schemes. Newer models cover the image with 32 by 32 pixel patches, shrink it if it exceeds a per-model and per-detail patch budget, then multiply the patch count by a model multiplier such as 1.2. Older models like gpt-4o and gpt-4.1 use 512 pixel tiles, with 85 base tokens plus 170 tokens per tile at high detail.

Q: What does Gemini's media_resolution setting cost per image?

For Gemini 4 and Gemini 3 models, Google's media resolution page lists about 280 tokens per image at low, 560 at medium, 1120 at high, and 2240 at ultra_high, which is only available per content item. The unspecified default for images is also about 1120, so if you never set media_resolution you are effectively paying for the high setting on every image.

Q: Does sending a bigger image always cost more tokens?

Only up to the vendor's cap. Claude downscales anything above its tier limit, so a 12 megapixel phone photo costs the cap rather than its full size; the waste is bandwidth and latency. Gemini 3 and 4 charge a fixed bucket regardless of size. OpenAI patch models shrink images to a patch budget. Resize on the client for speed, not for tokens above the cap.

Q: How do I choose the right image resolution for OCR?

Measure it on your own documents. Take about fifty real samples, run extraction at three or four resolutions, score the results against ground truth, and pick the cheapest resolution within a point or two of the best. Google notes PDF quality typically saturates at its medium setting. Crop to the region containing the text before you resize, so detail is kept where it matters.

Q: Why did my vision API bill go up when nothing changed in my code?

The usual cause is a model upgrade behind a floating alias. Newer models often accept larger images before downscaling, for example Claude's high-resolution tier, so the same 1920 by 1080 screenshot goes from about 1,560 to about 2,691 tokens. Pin model versions for vision features and log estimated versus billed tokens so a tokenizer change raises an alert.

Q: When is building an image preprocessing layer not worth it?

When volume is low. A few hundred images a day at around 1,500 tokens each costs very little, and an image preparation service is engineering time you will not earn back. The layer pays off for document pipelines handling tens of thousands of pages, computer-use agents that screenshot every step, and consumer apps where every user uploads photos.

Satyam Kumar

Zurück zum Blog

architecture

The Resize Is the Price: What Vision Tokens Really Cost in 2026, and Where to Cut Them

By Satyam KumarOctober 31, 202613 min read

vision token cost image token calculation multimodal llm pricing claude vision tokens claude high resolution images openai image tokens openai detail parameter gemini media_resolution gemini image tokens image preprocessing pipeline llm cost optimization computer use screenshot cost document ocr llm cost multimodal rag cost ai finops vision api budget

The Resize Is the Price: What Vision Tokens Really Cost in 2026, and Where to Cut Them

Frequently Asked Questions

Artikel teilen

Twitter LinkedIn WhatsApp

Satyam Kumar

Founder & AI Architect, AppScale LLP

KI & Cloud Architekt. Ich helfe Teams, Systeme zu bauen, die auf Millionen skalieren.

LinkedIn GitHub

The Resize Is the Price: What Vision Tokens Really Cost in 2026, and Where to Cut Them

Frequently Asked Questions

Artikel teilen

Comments

Leave a comment