Vision API pricing pages rarely say “per image”. They list a price per million tokens, and the image becomes tokens. This post turns those token prices into cost per image and per document page for the three vision models on BroadNet Inference, and shows which settings move the number.
How is an image billed in an LLM API?
A vision language model splits the image into patches and turns them into tokens. You pay for those tokens at the input price, together with your text prompt. The answer is billed as output. So:
cost per image ≈ (image tokens + prompt tokens) × input price / 1M
+ answer tokens × output price / 1M
There is no separate image fee on BroadNet. The image token count depends on the model’s image encoder and on resolution, so a 4,000-pixel phone photo costs more than the same photo resized.
How many tokens is one image?
It depends on the model family:
- Gemma 4 works to a fixed token budget per image: “70, 140, 280, 560, or 1120 tokens”, according to Google’s Gemma docs. Lower budgets downscale the image and run faster; Google says higher budgets keep fine detail and recommends them for the best accuracy on visual tasks.
- Qwen3-VL has a compression ratio of 32, so roughly one token per 32 × 32 pixel block, and its README shows how to set the visual token budget per image to a range such as 256 to 1,280. A 1,024 × 1,024 image is about 32 × 32 = 1,024 tokens.
BroadNet does not publish the image settings for its deployments, so treat these as orders of magnitude. Send one of your real images and read usage.prompt_tokens, then subtract the text-only count. That is your image token cost. The image token calculator applies each provider’s documented rule to any image size.
What does each vision model cost?
BroadNet Inference prices per 1M tokens, synced 2026-11-14:
| Model | Input | Output | Context | Reasons first | Starting credit |
|---|---|---|---|---|---|
| Gemma 4 VL 26B | $0.10 | $0.22 | 32,768 | Not listed | Yes |
| Qwen3.8 27B | $0.175 | $2.17 | 262,144 | Yes | After top-up |
| DeepSeek V4.1 Flash | $0.175 | $0.52 | 1,048,576 | Yes | After top-up |
What do 1,000 images cost?
Assume each image plus its short prompt comes to about 1,000 input tokens. Then 1,000 images is 1M input tokens, and the input cost per 1,000 images equals the price per 1M:
| Model | Input cost per 1,000 images | Output cost per 1M answer tokens |
|---|---|---|
| Gemma 4 VL 26B | $0.10 | $0.22 |
| Qwen3.8 27B | $0.175 | $2.17 |
| DeepSeek V4.1 Flash | $0.175 | $0.52 |
Run your own volumes through the free LLM API cost calculator. Two things change the ranking:
- Answer length. A one-word label is cheap everywhere. A 300-token JSON extract on a reasoning model can carry several hundred more reasoning tokens, billed at the output price. Qwen3.8 27B’s output price ($2.17 per 1M) is several times DeepSeek V4.1 Flash’s ($0.52), so for long answers DeepSeek costs less.
- Image size. If your images come to 2,000 tokens each, double the input column. If Gemma’s 280-token budget is enough for your task, the input cost falls to roughly a quarter.
What does one document page cost?
A scanned A4 page rendered at about 1,024 × 1,024 pixels lands near 1,000 tokens on a Qwen-style encoder. So a rough rule is 1,000 pages ≈ 1M input tokens, before the prompt and the answer. Small print may need more resolution, and more tokens.
Practical page tips:
- Send text PDFs as extracted text, not images. Text is far fewer tokens than a picture of text.
- Crop to the region you need, such as the totals box on an invoice.
- Send one page per request when you need per-page fields, so a failure does not cost the whole document.
Is BroadNet the cheapest vision API?
Not on every line. Here is the same family of open vision models at DeepInfra, from its pricing page on 14 November 2026, per 1M tokens:
| Model at DeepInfra | Input | Output |
|---|---|---|
| gemma-4-26B-A4B-it | USD 0.07 | USD 0.34 |
| Qwen3-VL-30B-A3B-Instruct | USD 0.15 | USD 0.60 |
| Qwen3-VL-235B-A22B-Instruct | USD 0.20 | USD 0.88 |
DeepInfra’s Gemma 4 input price is below BroadNet’s $0.10, and its output price is above BroadNet’s $0.22. So for image classification with one-word answers DeepInfra costs less; for long JSON extracts BroadNet can cost less. Compare more hosts on the Gemma 4 26B A4B price page and the Qwen3-VL 235B price page, or read the DeepInfra alternative page. The cheapest LLM APIs ranking and the LLM API price index put these prices in market context.
Which vision model should you choose?
- Lowest BroadNet price, simple tasks: Gemma 4 VL 26B. Labels, captions, “is there a signature?”. It works on the $5 starting credit. Its 32,768-token context limits how many images fit in one request.
- Arabic documents and screenshots: Qwen3.8 27B, the model BroadNet tested in Arabic. Read Arabic invoices with vision and choosing an LLM for Arabic.
- Long outputs, many images per request, or very high volume: DeepSeek V4.1 Flash. Its 1,048,576-token context fits many pages at once, and on a flat plan images are part of the unlimited use.
Side-by-side pages: Gemma 4 VL 26B vs Qwen3.8 27B and DeepSeek V4.1 Flash vs Gemma 4 VL 26B.
When does a flat plan beat per-image pricing?
A Full Moon session costs $60 a month for unlimited DeepSeek V4.1 Flash, one request at a time. At about 1,000 tokens per image, every 1,000 images cost $0.175 in input on DeepSeek V4.1 Flash, so input alone reaches the plan price at $60 ÷ $0.175 thousand images a month, before answers. If a single worker processes images all day, measure its throughput on a one-hour Meteor session ($1) and compare, or use the flat-rate vs per-token calculator. The full method is in the break-even math.
How can you cut vision costs?
- Resize to the smallest resolution that still reads correctly. Test three sizes on 20 samples.
- Ask for short JSON with named keys, not prose.
- Set
reasoning_efforttolowon reasoning models for simple extraction, and keepmax_tokensat 1,000 or more so the answer is not cut off. Why answers come back empty. - Use the correct MIME type in the data URL (
image/pngfor PNG). - Log
prompt_tokensper image to catch oversized uploads.
Where BroadNet Inference fits
BroadNet offers three vision models on one key, from Gemma per token to a flat plan, with no per-image fee and zero data retention for the images you send. It suits invoices, receipts, screenshots and Arabic documents. It is not the cheapest host for input-heavy work, as the DeepInfra table shows, and it does not offer image generation, only image understanding. For the extraction pipeline itself, see extract invoice data to JSON with a vision LLM. Start on the $5 starting credit.
