GPT Image 2.6 pricing: what one edit really costs on fal, measured

王立戈 — Wed, 11 Nov 2026 01:08:07 +0000

I run a small photo editor that calls three image models: Google's Nano Banana 2.1 through the Gemini API, and OpenAI's GPT Image 2.5 and 2.6 through fal.ai. Before I set a price I run each model on the same photos and read the bill, because per-image estimates rarely match what an edit actually costs. GPT Image 2.6 came out on November 4; these numbers are from November 5 and 6.

GPT Image 2.6, per image

Flare endpoint (openai/gpt-image-2.6/flare/edit), one 4:3 input photo, output at the same 4:3 shape. These are whole requests: the input photo, the prompt and any reasoning tokens are included.

`quality`	1K	2K	4K
`medium`	$0.041	$0.075	$0.119
`high`	$0.048	$0.089	$0.142
`max`	$0.164	$0.324	$0.540

The surprise is the medium row. On GPT Image 2.5 the same settings cost $0.019, $0.029 and $0.042, so 2.6 at medium is two to three times the price. high and max cost exactly what they did on 2.5.

Each extra input image (a second photo, or a mask sent as an image) added about $0.0085, whatever the quality.

fal now bills 2.6 by tokens ($30 per million image-output tokens, $8 per million image-input tokens) and lists an estimate per output size. One thing in that table is easy to miss: square images cost more than wider or taller ones with more pixels. At medium, the image output for 1024 × 1024 is $0.043, for 1024 × 1536 it is $0.034. If you don't need a square, don't ask for one.

GIFs

2.6 can return an animated GIF, and fal's edit endpoint takes a photo and gives one back with output_format: "gif". One photo to a 576 × 432 GIF, 9 frames:

`quality`	cost
`low`	$0.035
`medium`	$0.054
`high`	$0.064

A GIF has 256 colors, and I couldn't see a difference between the three. Bigger frames (768 × 576) cost $0.073 and gave a 2.5 MB file that didn't look better. Setting reasoning_effort to low didn't change the price. I settled on medium, about 512 × 512 pixels per frame, and the default 9 frames of 200 ms, which is enough for one small motion.

Compared with Nano Banana 2.1

For Nano Banana 2.1 I converted the token counts in each response's usageMetadata at the official prices (October 9, a 4:3 input): 1K $0.044 (about 25 s), 2K $0.057 (about 29 s), 4K $0.121 (about 44 s). An extra reference image is about 1,120 input tokens, roughly $0.0017.

So at 1K, GPT Image 2.6 at medium ($0.041) and Nano Banana 2.1 ($0.044) now cost about the same. At 2K Nano Banana is cheaper ($0.057 against $0.075), and extra images are five times cheaper.

Three things worth knowing before you ship it

1. How the mask reaches the model. When I first wired up GPT Image 2.5, I passed the brushed area as fal's mask_url. Edits inside the mask came back unrelated to what was there: color and texture changes redrew the whole shirt or face. Sending the mask as one more input image, with a sentence after the prompt saying the last image is a mask and only its white area should change, fixed it. I send masks to 2.6 the same way and haven't gone back to mask_url.

2. 2.6 adds real brand names. I ran 48 edit prompts on Flare. In 3 results it printed real brand names that weren't in the photo: book titles on a living-room shelf, a logo on a gaming headset, a toy maker's logo on a figure box. Of 10 prompts I also ran on Sunburst, 1 had a real brand on a speaker. If the results end up in product listings, check the props.

3. Output size rules. GPT Image wants both sides a multiple of 16, the longest side at most 3840, an aspect ratio of at most 3:1 and at most 3840 × 2160 pixels in total. I compute width and height from a target pixel count and the photo's aspect ratio, then round down to 16, so a request never goes over.

How I price it

My rule is that the credits for an edit, at my cheapest bulk rate, cover at least twice the measured cost. With 2.6's medium price, a 1K edit on 2.6 is 6 credits where 2.5 was 3. Measure before you price: run a few edits on your own photos and read the bill; Gemini also puts the token counts in every response.

The 48 test edits are side by side, Nano Banana 2.1 against GPT Image 2.6, here: https://yepshot.com/nano-banana-2-1-vs-gpt-image-2-6?utm_source=devto&utm_medium=social. The prompts and results are open, CC BY 4.0: https://github.com/lygerone/awesome-gpt-image-2-6-prompts