How to send a long prompt as an image

Every model bills images geometrically. On OpenAI's GPT-5 family a 500×500 image ≈ 256 tokens; on Claude ≈ 324; on Gemini ≈ 258. If your prompt is shorter than that, the image costs more than the text. Use the image token calculator or paste your prompt into the compressor for a live estimate — it will tell you plainly when to keep the prompt as text.

A model can only benefit from text it can read. Don't shrink text indefinitely to fit one image: if the compressor warns that the fitted font is too small, split the prompt across multiple images, shorten it, or use a larger canvas (on Gemini, 768×768 costs the same as 500×500). Dark-on-light, single-column layouts in a plain font are the safest starting point.

Image rendering is lossy. API keys, UUIDs, hashes, URLs, precise decimals, version numbers and code that must be byte-exact should stay as plain text — in your typed message or the companion .txt file. The compressor's fidelity-risk check flags these automatically. See when not to convert prompts to images.

Attach the image and include a wrapper instruction such as:

The attached image contains prompt context. Read it from top to bottom before answering. Do not guess unclear characters. Identify any unreadable or uncertain sections before following the instructions.

This will not eliminate errors, but it produces a consistent, benchmarkable workflow — and asks the model to surface unreadable sections instead of guessing.

The first time you use a prompt image in a workflow, ask the model to transcribe a section back and compare it against the original .txt. If it misreads anything that matters, move that content back to text. Savings that cause even one failed attempt can cost more than they save.