Language

Vision-language & image understanding

Models that read photos, screenshots, charts and documents and answer in text.

What it is

Vision-language models (VLMs) take an image, a screenshot, a chart or a scanned page together with a question and reply in text or JSON. They describe photos, read charts, extract fields from forms, check UI screenshots, and ground answers to regions of an image. Many general LLMs now accept images; this field compares the image-understanding side specifically.

How pricing works

Images are converted to tokens (usually 250–2,000 tokens for a ~1 megapixel image, depending on the provider's tiling) and billed at the model's input-token price; the text answer is billed as output. We normalise to the input cost of 1,000 images of about 1 MP using a ~1,500-token estimate, so output text and your prompt come on top.

We compare everything in one unit: USD per 1K images (≈1 MP, input only).

Typical uses

Describing and tagging product or user photos
Chart, diagram and screenshot analysis
Visual QA for support and accessibility (alt text)
Extracting data from invoices and forms when OCR alone is not enough
UI testing and computer-use agents that need to see the screen

What to compare

Accuracy on visual QA, chart and document benchmarks
Maximum images per request and image resolution
Grounding (bounding boxes / pointing)
Video-frame support
Latency per image

Watch out for

Tokens per image differ widely between providers, so a lower per-token price does not always mean a lower per-image price.
High-resolution or tiled images can cost several times a low-detail image; set the detail level explicitly.
VLMs can hallucinate small text or numbers; for exact digits use a dedicated OCR/document API and compare.

Capabilities explained

Text input price · $ / 1M tokens: Price of the text part of your prompt. Images are priced separately in the headline number.
Max images per request: How many images can be sent in one call; matters for comparing pages or frames.
Grounding / bounding boxes: Can return coordinates of objects or text regions instead of only describing them.
Video input: Accepts video clips or many frames, not only still images.
Document & chart understanding: Tuned for PDFs, tables, charts and screenshots rather than only natural photos.
Open weights: The model weights are downloadable, so you can self-host it or switch hosting providers without changing models.

Glossary

Image tokens: The images are split into tiles and converted to tokens; larger images use more tiles and cost more.
Grounding: Returning where in the image something is (boxes or points), not only describing it.
Detail level: A setting that trades image resolution (and cost) against accuracy on small details.

All 4 APIs, cheapest first

Free to browse

Tick 2 to 4 APIs to compare them side by side.

API	$ / 1K images	Max images per request	Grounding / bounding boxes	Video input	Quality	Price data
GLM-4.6V-Flash Z.ai (Zhipu GLM)	Free	10	—	No	–	Verified Oct 6, 2026
Qwen-VL (Model Studio) Alibaba Cloud (Qwen)	$0.60	—	Yes	Yes	–	Unverified sample
Gemini Flash (image input) Google (Gemini / Vertex)	$0.75	3,600	Yes	Yes	–	Unverified sample
Grok 4.3 (image input) xAI (Grok)	$1.90	—	—	No	–	Unverified sample