Measuring the Per-Image Cost of GPT-Image 2.6: Does Higher Quality Improve the Image?

GPT-Image 2.6 was released a week ago. How much does it actually cost to make one image, and do the images change as much as the quality level suggests?

The official documentation lists token prices, but not per-image prices for medium and higher quality levels in GPT-Image 2.5 and 2.6. So I generated 69 images while changing the model and quality level, then calculated the cost from the token counts in the responses. I also cover the process and results in the video.

How I compared the models

I compared gpt-image-2, GPT-Image 2.5 Flare and Sunburst, and GPT-Image 2.6 Flare and Sunburst. The prompts were for a jazz poster with text, a product photo of an earbud case with a “UNION” label, and a winter-night tteokbokki street stall with a sign in Korean.

I used five quality levels: low, medium, high, xhigh, and max. I requested 1024×1024 images and generated one image for each combination, for a total of 69 images. The total cost was $5.63. gpt-image-2 does not support xhigh or max, so those requests returned a 400 error.

The price of one 1024×1024 image

I calculated the price using $30 per 1 million output tokens and $5 per 1 million input text tokens. The prices below are for one image. For GPT-Image 2.5 and 2.6, Flare and Sunburst had the same price at every quality level.

All 69 images, by-quality view and 1:1 zoom: https://bgprogramming.com/dev/image-model-comparison

For 2.6, the price goes from $0.025 at low to $0.211 at max, a difference of about 8.5 times. For 2.5, the price goes from $0.006 at low to $0.211 at max, or 35 times. Compared with 2.5 low, 2.6 low costs 4 times as much.

Do images change when you raise the quality level?

I compared the same street stall image from low to max with 2.6 Sunburst. The composition stayed almost the same as I raised the quality level. I saw small differences in textures such as knit fabric, steam, and wood grain. They were easier to see when I zoomed in on the original at 1:1. At thumbnail size, it was hard to tell them apart.

Even at low, the English and Korean text was accurate, including the small Korean menu board. The price goes up, but with the three prompts I tested and at thumbnail size, the differences between quality levels were not large.

Model differences mattered more than quality levels

For the jazz poster, 2.5 added small text I had not requested. One example was “GOOD MUSIC BRINGS US TOGETHER.” Raising the quality level did not fix this, and the text was not present in 2.6. Even at the same low level, the scene from 2.6 was more crowded.

The price difference also depended on the model. At high, gpt-image-2 cost $0.211 and 2.6 cost $0.053. Generation time also differed: gpt-image-2 high took an average of 105 seconds, while 2.6 Flare high took 10 seconds.

Flare and Sunburst: same price, different speed

2.6 Flare and Sunburst had the same price at every quality level. In this comparison, the speed difference was clear. The times below are averages across the three prompts.

Sunburst was 2–5 times slower than Flare. Flare low took 7–8 seconds in all three runs. That does not mean Flare is a lower-quality model. Text accuracy was the same at low.

OpenAI describes Sunburst’s strengths as medium and higher quality, and consistency during iterative editing. In my comparison, the difference between medium and high was not clear, and I did not measure iterative editing. I might consider Flare when users are waiting or when I need to generate several drafts, and Sunburst when I need to revise the same image multiple times or keep a character consistent. That is my judgment based on these measurements.

Token counts looked fixed — limitations of this comparison

For the same model and quality level, the output token count was the same across different prompts. For example, 2.6 low used 824 tokens, and medium used 1,449 tokens. The counts were the same for a simple poster and a night scene with people and signs.

This is based on only three prompts, so I cannot say it is a general rule. But in this test, output token counts seemed to be fixed by resolution and quality level rather than by image complexity.

I also generated only one image for each combination, so the composition could change if I generated it again. I used only three types of prompts, and I did not test iterative editing. Please treat the prices here as results from this test of one 1024×1024 image.

What I changed afterward

My channel thumbnail tool was set to gpt-image-2 high. After this comparison, I decided to use GPT-Image 2.5 Sunburst at low.

When I made this video’s thumbnail with gpt-image-2, “8.5 times,” which I had written in the first line of the prompt as an explanation, appeared as text in the image. The phrase “The highest quality costs 8.5 times more!” (originally in Korean) appeared in two places, even though I had not requested it. I removed that sentence, added “No other text,” and generated the image again.

Three things to check

First, compare images at the size you will actually use, starting with the lowest quality level. Raise it only if the results are not good enough.

Second, check the cost per image instead of looking only at the token price. You can calculate the cost by recording the token counts in the response.

Third, when a new model comes out, compare it using the prompts you already use.

When generating images, which matters most to you: speed, price, or quality?

This article summarizes my Korean YouTube video. It was translated with AI and reviewed by me. #ABotWroteThis

Original video (Korean, with English subtitles): https://youtu.be/Cz281z_MplA

Previous video: Don't Use an AI-Generated DB Schema As-Is