How are image tokens calculated for GPT-4o?

Low detail uses a fixed base of 85 tokens per image. High detail scales the image to fit within 2048×2048, then to a 768px short side, splits it into 512×512 tiles, and charges 170 tokens per tile plus the 85-token base. More pixels means more tiles and more tokens.

Does adding an output prompt change the cost?

This estimator focuses on the image input tokens, which dominate vision cost. Any text prompt and the model's text answer are billed separately at the usual input and output rates — add those with a standard LLM cost calculator if they are significant.

Why is low detail so much cheaper?

Low detail collapses any image to a single fixed-cost representation (85 tokens for GPT-4o), ignoring resolution. It is ideal for classification or presence checks where fine detail is not needed, and can be 5-20× cheaper per image.

Is my data sent anywhere?

No. Everything is computed in your browser. Nothing you enter is uploaded, stored or logged.

Vision API Batch Cost Estimator

Estimate vision API cost before you run the batch

Vision models bill images as tokens, and the count depends on resolution and detail level — not a flat per-image fee. A high-detail 4K image can cost many times a low-detail thumbnail. This estimator computes per-image tokens the way GPT-4o does, converts to dollars, and scales to batches from 100 to 1 million images so you can budget a vision job accurately.

How it works

For GPT-4o-style tiling, a low-detail image is a flat 85 tokens. A high-detail image is resized to fit 2048×2048, then its short side is scaled to 768px, split into 512×512 tiles, and billed as:

tiles  = ceil(scaled_w / 512) × ceil(scaled_h / 512)
tokens = 85 + 170 × tiles

The per-image dollar cost is tokens / 1,000,000 × input_price, and the batch cost simply multiplies by your image count. Gemini uses a flat per-image token charge in this estimator, which is how its vision pricing is commonly modelled.

Tips to cut vision cost

Use low detail for classification. If you only need to know what is in the image, low detail is far cheaper and usually accurate enough.
Downscale before sending. There is no benefit to sending a 6000px image if the model resizes it anyway — resize client-side and skip wasted tiles.
Crop to the region of interest. Sending only the relevant part of an image cuts tiles and tokens directly.
Batch with the async API. For large jobs, the batch endpoint is often cheaper than real-time calls and avoids rate-limit churn.