How Many Tokens Is an Image? Vision Model Cost
An image is converted into tokens before the model reads it, and the count depends on the encoder and resolution. If an image becomes 1,000 tokens and input costs $2 per 1M, each image costs $0.002, so 100k images a month costs about $200.
Last updated: October 2026
Key takeaways
- Image token counts vary widely by encoder and resolution, so there is no single answer to how many tokens an image is.
- Image tokens are billed as input before any output is generated.
- A 10x spread in tokens per image becomes a 10x spread in your input bill.
- Budget vision workloads per image, with a low, typical and high resolution scenario.
- Downscale, crop or tile images deliberately instead of sending full-resolution uploads.
How many tokens is an image?
There is no fixed number. Spheron explains that image tokenization varies wildly by encoder and resolution, and that the inflation happens before the model generates a single output token. A small thumbnail may become a few hundred tokens while a large detailed photo can become thousands. Check your provider's documentation for the exact rule, because each vendor tiles and counts images differently. Our post on tokenizer inflation shows how the same text can already differ across vendors.
Why do image tokens cost more than text?
They do not necessarily cost more per token. They cost more per piece of content. A paragraph of text might be 200 tokens. A single image can be 1,000 or more, so a prompt with one photo can cost several times as much as the text around it. The price per token is the same input rate, but the count is far larger and harder to predict.
What does a vision workload actually cost?
Say you run an expense-receipt reader that processes 100k images a month at an illustrative input rate of $2 per 1M tokens. The token counts below are assumptions to show the spread, not measured values for any model.
| Resolution scenario (illustrative) | Tokens per image | Cost per image | Monthly cost at 100k images | |---|---|---|---| | Low, downscaled | 300 | $0.0006 | $60 | | Typical | 1,000 | $0.0020 | $200 | | High, full resolution | 3,000 | $0.0060 | $600 |
The same receipts cost 10x more at the high setting. If you charge your customer a flat $0.01 per scan, your gross margin on inference alone is 94 percent at the low setting and 40 percent at the high setting, before output tokens and retries.
How do you control image token cost?
Four moves usually work, and you can test each one on a sample set:
- 1Downscale images to the smallest size that keeps your task accurate, such as readable text on a receipt.
- 2Crop to the region that matters instead of sending the full frame.
- 3Cap the number of images per request and per user per day.
- 4Cache results for repeated images so you never pay twice.
Agents that look at screenshots are especially exposed, because every step can add an image. See how that plays out in our post on computer-use agent costs.
Should you price vision features per image?
If your cost per image varies 10x, a flat monthly plan with unlimited uploads is risky. Consider a credit model where one scan equals one credit at a standard resolution and larger files cost more credits. That keeps your margin predictable without surprising users. Compare current provider rates on the LLM pricing hub before you set credit values.
Credit to the original explainer: Spheron's Cost Per Token Vision Model: Image Tokens Cost More, which walks through why vision cost math differs from text.
Takeaway: measure tokens per image on your real files, then price the feature on the high end of your range.
To model your own image volume, try the LLM cost calculator.
Frequently asked questions
How many tokens is an image for an LLM?
It depends on the model's encoder and the image resolution. Counts range from a few hundred to several thousand tokens per image, so check your provider's documentation and measure a sample.
What is the cost per image for a vision language model?
Multiply tokens per image by the input price per token. For example, 1,000 tokens at $2 per 1M tokens is $0.002 per image. Add output tokens for the model's response.
Why does a vision model cost more than a text model?
Images expand into many tokens before the model reads them, so one image can outweigh pages of text. The per-token rate may be the same, but the token count is much larger.
How can I reduce vision model costs?
Downscale or crop images, limit images per request, and cache repeated results. Test accuracy at each setting so you cut cost without hurting quality.
FAQ schema (JSON-LD)
Place this inside a <script type="application/ld+json"> tag in the page head.
More from the blog
LLM EconomicsJevons Paradox in AI: Why Cheaper Tokens Mean Bigger Bills
Jevons paradox in AI means that when tokens get cheaper, people use so many more that total spend rises. If price per task halves from $0.10 to $0.05, volume must double just to hold spend flat, and a tripling lifts spend 50 percent.
LLM EconomicsOpenRouter Pricing: What a Gateway Really Costs
OpenRouter pricing is the underlying model rate plus whatever platform fee the gateway adds, so the real question is net cost. Say a 5 percent fee on $10,000 of monthly tokens adds $500, and routing 15 percent of spend to cheaper models saves $1,500, a net gain of $1,000.
LLM EconomicsMoE Cost per Token: Why Active Parameters Rule
For a mixture-of-experts model, compute per token follows active parameters, while the GPUs you must rent follow total parameters. A model with 40B active out of 400B total can be cheaper per token than a dense 70B at high utilization, and more expensive when traffic is thin.
Pricing math, in your inbox.
One short note a week on AI pricing, token economics, and margin. No spam, unsubscribe anytime.