Does a longer prompt cost more on GPT Image 2.5? Text token math

Yes, on GPT Image 2.5 a longer prompt costs more, but not by much per word. Fal lists text input at $5 per million tokens, so a 2,000-token brief adds about $0.01 to an image. That is about the same as the whole output estimate for a medium 1024x1024 image ($0.01317). For a short prompt the text is a rounding error; for a long style brief sent on every call, it can double a cheap draft.

The arithmetic

Fal's GPT Image 2.5 Flare page lists text tokens at $5.00 per million input, $1.25 cached, $10.00 output, and image tokens at $8.00 input, $2.00 cached, $30.00 output. It also says longer prompts increase the cost. The Sume docs use the same Fal token rates and say input token counts are estimates. Token counts for English run roughly 1.3 tokens per word, which is our approximation, not a figure from either vendor.

Where it starts to matter

Compare against output: at 1024x1024 the output estimate is $0.00588 for low, $0.01317 for medium, $0.05268 for high. A 500-word brief adds $0.00325 which is half of a low image and a quarter of a medium one. On high or above it vanishes.

The pattern is the one that bites: a shared 500-word brand brief pasted into 1,000 draft calls adds $3.25. Moving the stable part of the brief out of the prompt and into a reference image is not free either, because input image tokens are billed at $8 per million.

Ways to keep it small

• Put the subject first and the style rules after it, and cut anything the model already does by default.
• Reuse one reference image for style instead of a paragraph describing it; read the estimate for the reference tokens on a test call.
• Draft at low, and write the long brief only for the final pass.
• Keep a house style as a short string in your wrapper, not a pasted document.

Verify on your own account

Cached input is cheaper on Fal's list ($1.25 text, $2.00 image), but whether Sume applies a cached rate to your repeated prompts is not something the Image API docs promise, so do not budget on it. Send the same call with a short and a long prompt at n: 1 and subtract the two usage.cost values. That difference is your real price per extra word on Sume.