Pricing

• Per-model rate card. Each model exposes a pricing object with USD-per-1M rates for prompt, completion, input_cache_read, input_cache_write, and image.
• Per-request cost. usage.cost on a chat completion, streaming or not, is the actual metered charge for that request, in USD rounded to six decimals, the resolution the ledger bills at. Embeddings and transcriptions carry usage.cost too: a transcription against input_audio_seconds, on the json and verbose_json formats. A cost that rounds to 0.0 at six decimals is a real zero, not a missing value.

​

How billing works

• Token rates are USD per 1M tokens. Input is prompt tokens, Cached input is input_cache_read, and Output is completion tokens. A request pays prompt on the uncached prompt tokens, input_cache_read on the cached ones, and completion on the output. The one model where prompt is not a token rate is nvidia/parakeet-tdt-0.6b-v3, where it is per audio second.
• Documents bill per page. The Gateway rasterizes a PDF and runs one inference per page. It sums the per-page token counts, so the charge is linear in page count.
• input_cache_write is $0.00 across all models today.
• Which rate applies is decided by the model’s own rates, not by the method: a model bills per input when its prompt rate is $0.00 and its image rate is not, and per token otherwise. Per input means once per image, or once per page on a PDF. facebook/sam3.1 is the only served model on the per-input side today. See Segmentation and tracking.
• A request carrying a video bills the video rate once per video whenever that rate is set, whichever mode the model is otherwise in. A clip contributes no images, and a flat per-image token count says nothing about its length.
• Several token-billed models publish a non-zero image rate that they are not charged on, zai-org/glm-ocr, deepseek-ai/deepseek-ocr-2 and paddlepaddle/paddleocr-vl-1.6 among them. Read usage.cost rather than multiplying the image rate yourself.
• Neither method (ocr, markdown, detect, parse_layout, chat) nor response_format (text or JSON) changes the rate. You pay the rate of whichever model serves the request.
• Non-generative vision models emit no text. They report a flat prompt-token count per image instead, so the token rate gives a fixed price per image.

​

What 1M tokens buys

• ~3,000 image captions at ~200 input + 130 output tokens / image
• Markdown from a ~2K-page slide deck with figures at ~500 output tokens / page
• ~4,000 visual questions at ~200 input + 50 output tokens / question
• ~2,000 text passages embedded at ~512 tokens / passage
• Markdown from ~1K pages of legal docs with dense text at ~1,000 output tokens / page

​

Chat and document models

Prices in USD per 1M tokens

​

Pose estimation

Prices in USD per 1M tokens

​

Segmentation and tracking

$1.00 per 1,000 images. A video is billed once, whole, however many frames are sampled.

​

Frontier models

Prices in USD per 1M tokens

​

Embeddings

Prices in USD per 1M tokens

​

Transcription

About 16.7 hours of audio per dollar

"usage": {
  "type": "duration",
  "input_audio_seconds": 252.726,
  "output_tokens": 581,
  "cost": 0.004213
}

​

How to get model pricing

import httpx
from urllib.parse import quote

model_id = quote("paddlepaddle/paddleocr-vl-1.6", safe="")
response = httpx.get(
    f"https://gateway.vlm.run/v1/models/{model_id}",
    headers={"Authorization": "Bearer <VLMRUN_API_KEY>"},
)
print(response.json()["pricing"])

const modelId = encodeURIComponent("paddlepaddle/paddleocr-vl-1.6");
const response = await fetch(`https://gateway.vlm.run/v1/models/${modelId}`, {
  headers: { Authorization: `Bearer ${process.env.VLMRUN_API_KEY}` },
});
const { pricing } = await response.json();
console.log(pricing);

curl "https://gateway.vlm.run/v1/models/paddlepaddle%2Fpaddleocr-vl-1.6" \
  -H "Authorization: Bearer $VLMRUN_API_KEY"