Same uploaded image, same model. Changing the detail setting to Low increases this image's input cost by about 64%. These are calculated estimates using OpenAI's November 16, 2026 rules.
OpenAI image tokens come with a surprisingly unhelpful setting: Low can cost more than High. On GPT-5.4, a 2048 by 2048 image uses an estimated 4,916 tokens on Low, against 3,000 on High. I'd expect the cheaper option to be obvious from its name. Apparently that's asking too much.
There's a second trap when you switch models. Leave image detail on Auto, and a 4K screenshot can go from 2,765 image tokens on GPT-5.4 to 9,792 on GPT-5.6 Sol. The default starts preserving more pixels. Your request doesn't need a new line of code for that part of its cost to change.
I reproduced the current sizing rules across desktop, phone, tablet, square, and long webpage dimensions, then checked the examples below in OpenAI's image input calculator (opens in new tab). When I send a screenshot by API, I want to know how many pixels reach the model under the setting I've chosen.
Why GPT-5.4 can charge more for Low image detail
For these models, OpenAI counts image patches measuring 32 by 32 pixels. A partial patch still counts. The model's preprocessing rules determine which dimensions get counted.
Take that square upload. GPT-5.4 Low leaves it at 2048 by 2048: 64 patches across and 64 down, or 4,096 patches. Multiply by 1.2 and round up, and you get 4,916 image tokens.
High reduces the same upload to 1600 by 1600. That fits its 2,500 patch budget, giving 3,000 tokens after the same multiplier. High is cheaper because it sends a smaller image through this part of the pipeline. OpenAI's model sizing table (opens in new tab) explicitly warns that Low can use more tokens on GPT-5.4, including its mini and nano variants.
Low keeps the square at 2048 pixels on each side. The calculator's warning is immediately below the detail choices.
High shrinks it to 1600 by 1600. That's why its estimate is lower.
At the listed standard input rate, 10,000 of these square images would cost about $123 on Low versus $75 on High, for image input alone. That makes a poorly named option worth checking in a document or screenshot pipeline.
The shape matters, though. A 1920 by 1080 image stays below High's patch budget, so both settings use 2,448 tokens. A 4K widescreen image also ends up at the same size under Low and High on GPT-5.4. You won't find the 64% gap on every upload.
| Uploaded dimensions | GPT-5.4 Low | GPT-5.4 High | Extra tokens on Low |
|---|---|---|---|
1920 x 1080 | 2,448 | 2,448 | 0 |
1290 x 2796 | 2,304 | 2,304 | 0 |
2732 x 2048 | 3,687 | 2,942 | 745 |
2048 x 2048 | 4,916 | 3,000 | 1,916 |
3840 x 2160 | 2,765 | 2,765 | 0 |
A square dashboard crop and a wide desktop screenshot can hit different branches of the same pricing rules. The square is where I'd check for this mistake first.
Auto changes meaning when you change the model
Auto is the default when you omit image detail. Its behavior depends on the model.
For GPT-5.4, Auto follows High's sizing. For GPT-5.5, GPT-5.6 Sol, and GPT-6 Astra, it follows Original. The same application code can therefore deliver substantially more image data after a model switch. OpenAI documents these defaults in its vision guide (opens in new tab).
Image tokens, rather than dollars: these models have different input prices. Each pair compares two settings on the same model.
On GPT-6 Astra, that 4K image stays at 3840 by 2160 on Auto or Original. High reduces it to 2104 by 1184, cutting its estimated image tokens from 9,792 to 2,931.
Original preserves this 4K image. Auto uses the same sizing behavior on GPT-6 Astra.
High uses about 70% fewer image tokens here, while keeping fewer pixels.
At Astra's listed standard input price, 10,000 such images work out to roughly $980 on Auto versus $290 on High. An occasional screenshot isn't going to bankrupt anyone. Repeated screen captures in a browser agent, visual monitoring job, or document workflow give the setting more chances to matter.
Original also has a purpose. If I'm asking for the location of a tiny button, reducing the image changes both the available detail and the coordinate system. I'd choose the resolution around the task, then check the cost. A cheaper upload that makes the task fail is wasted money.
Long webpage screenshots can shrink dramatically or hit a limit
A full webpage capture creates a different problem. Suppose it's 1440 pixels wide and 12,000 pixels tall.
On GPT-6 Astra, Auto preserves those dimensions and uses an estimated 20,250 image tokens. High fits the page into its patch budget, leaving 544 by 4533 pixels and 2,897 tokens. An 18 pixel label in the source would occupy about 7 pixels after that reduction.
GPT-5.6 Sol High has an additional 2048 pixel edge limit. It reduces the same long image to 246 by 2048, for just 615 tokens. That's a very cheap strip of a webpage. Before relying on it to read a dense settings panel, I'd crop out that panel instead.
Making the capture twice as tall exposes another boundary. At 1440 by 24,000, Astra Original needs 33,750 patches, above the per image limit of 30,000. The documented result is rejection; the API doesn't shrink this Original image to fit that patch limit.
A long capture can cross the input limit even though neither dimension reaches the separate 65,535 pixel edge limit.
For a long report, I'd send the relevant section with enough surrounding context to identify it. If the task needs several sections, separate crops can preserve more pixels around the text, though their token costs add up. Shrinking the whole report until it passes a size check doesn't tell me whether the information I need survived at a useful scale.
Check the uploaded dimensions before optimizing image costs
The image on your screen and the file reaching the API aren't necessarily the same size. A client can resize attachments before sending them. Calculate from the uploaded dimensions, rather than the camera resolution or the size of the source screenshot.
For an image workflow, I'd check these in order:
- Choose the area the task needs. A question about one dialog usually doesn't need a full desktop or twelve screens of webpage. Keep surrounding context when the question depends on it.
- Name the model and detail setting. Don't carry a Low or Auto assumption from one model to another. The examples here use specific API model identifiers, not ChatGPT subscription labels.
- Inspect the processed dimensions. A token estimate tells you cost; the resulting width and height help explain what you're giving up. Low on Astra reduces a 4K image to 512 by 288, whereas Low on GPT-5.4 follows a different path.
- Check the result on your own task. Use a few representative screenshots with answers you can verify. Compare task success and reported usage before rolling a cheaper setting across a pipeline.
These calculations cover image input at standard, uncached rates. Text, output, and other request charges come on top. For controlling the total bill, I'd also use spending limits and alerts, rather than relying on inexpensive individual requests.
How I calculated the examples
I implemented OpenAI's published pixel limits, patch budgets, rounding, and model multipliers, then
checked the examples against its calculator on November 16, 2026. The API model identifiers in the
main comparison are gpt-5.4, gpt-5.5, gpt-5.6-sol, and gpt-6-astra. I held the uploaded
dimensions fixed and changed the model or detail setting. These are image input estimates; I didn't
run a vision accuracy benchmark or compare invoices. OpenAI notes that billing rounding can differ
by one token.
Choose the model that fits the screenshot task
This is part of why I built Cumbersome: I want to choose the provider and model for the job, rather than move my work between separate AI subscriptions every time that choice changes. It's a native client for iPhone and Mac that connects through your own API keys, and Message Details shows the token usage the provider reports.
Cumbersome also prepares image attachments before sending them, currently capping their longest edge at 1,568 pixels. So a 4K source screenshot won't follow the full resolution example above unchanged. For Cumbersome, calculate from the prepared attachment's dimensions, then check the reported usage and the answer you got back.
Bless up! 🙏✨