Guide

When not to turn a prompt into an image

We build a visual prompt compressor — and these are the cases where we'll tell you not to use it. Trustworthy compression starts with knowing where it fails.

Keep your prompt as plain text when it is shorter than the model's image-token cost, when it contains exact identifiers such as API keys, UUIDs, hashes or precise numbers, when prompt caching already discounts the repeated prefix, or when a transcription mistake would have legal, financial or medical consequences. Image rendering is lossy and models can misread dense text silently.

1. The prompt is below break-even

A 500×500 image costs roughly 256 tokens on OpenAI's GPT-5 family, 324 on Claude and 258 on Gemini. If your prompt is shorter than that, the image is strictly worse: more tokens, plus fidelity risk. Check the calculator first.

2. It contains exact values

Never rely on a lossy image for content where one wrong character breaks something:

API keys, credentials, tokens
UUIDs, hashes, long identifiers
URLs, file paths, email addresses, phone numbers
Exact decimals, financial figures, version numbers
Code that must be byte-exact
Legal, financial or medical facts where a transcription mistake would be serious

Dense image recall can fail silently — the model answers confidently from a misread value. Keep these as plain text and use images only for prose context. The compressor detects these patterns and warns you before export.

3. Prompt caching already covers it

If you resend the same long prefix to the same API, provider prompt caching discounts it losslessly. Compare image tokens against the cached text price — not the full price — before concluding an image is cheaper. Caching wins whenever the context repeats verbatim; images win for portable, cross-tool context.

4. You pay a flat subscription

Consumer ChatGPT and Claude plans bill a flat subscription with usage limits, not per token. A prompt image may help you fit more context or stretch limits where image input is supported, but it does not directly reduce your monthly bill. Direct dollar savings apply to usage-metered workflows: APIs, agents, batch jobs.

5. The text would become microtext

If fitting your prompt into one image pushes the font below legibility, the model's reading accuracy collapses and any token savings are consumed by retries. Split across multiple pages or keep it as text — never ship unreadable microtext.

Last updated July 15, 2026 · Estimates are derived from provider documentation and can change — see methodology.