Vision Costs: GPT-4o, Claude, Gemini Image Pricing Compared

Image Tokenization Across Providers

Different providers tokenize images very differently. A typical 1024x1024 photograph:

Provider	Tokens per image (1024x1024)	Cost per 100 images (input)
GPT-4o (high detail)	~765	$0.19
GPT-4o (low detail)	85 (fixed)	$0.02
Claude Opus 4.7 (auto)	~1,600	$2.40
Claude Sonnet 4.6 (auto)	~1,600	$0.48
Gemini 2.5 Pro	~258	$0.032
Gemini 2.5 Flash	~258	$0.0077

How GPT-4o tokenizes images

GPT-4o uses a tile-based scheme:

• Low detail: fixed 85 tokens regardless of resolution.
• High detail: 85 base tokens + 170 per 512x512 tile. A 1024x1024 image needs 4 tiles → 85 + 4×170 = 765 tokens.

A 4096x4096 image (16 tiles) would be 85 + 16×170 = 2,805 tokens — about $0.007 input on GPT-4o. Always downscale to the smallest resolution that lets the model accomplish the task.

How Claude tokenizes images

Claude resizes any image to fit within 1568x1568 and then tokenizes at roughly (width × height) / 750 tokens. A 1024x1024 image ≈ 1,398 tokens; a wide 1920x1080 ≈ 2,765 tokens.

How Gemini tokenizes images

Gemini uses a fixed 258 tokens per image for images up to 384×384 (after resizing). Larger images are split into tiles of 768×768, each tile costing 258 tokens.

Practical guidance

• Receipt OCR / document parsing: Use GPT-4o low-detail (85 tokens flat). Even at 1080p the recognition is good for printed text.
• Image classification: Gemini Flash is dramatically cheaper. Claim quality on a benchmark first.
• Detailed visual reasoning (charts, diagrams, screenshots): GPT-4o high-detail or Claude. The extra tokens buy actual capability.
• Bulk image labeling: Gemini Flash by a wide margin. At ~$0.00008 per image you can label 1M images for $80.

The output side

Vision pricing usually focuses on input. The output (model's text response describing the image) is billed at the regular text output rate. A 200-token caption on GPT-4o is $0.002 — negligible compared to image input cost.