Methodology: how our estimates work

We estimate text tokens as ceil(characters / 4), the common English-language heuristic. Real tokenizers vary by model and language — code, non-Latin scripts and unusual formatting can tokenize very differently. Treat the number as a planning estimate, not a bill.

Per model family, from currently documented image-input billing:

• OpenAI / ChatGPT (GPT-5 family): ≈ one token per 32×32-pixel patch (rounded up per side). Example: 500×500 ≈ 256 tokens.
• Claude (Sonnet / Opus / Fable): ≈ one token per 28×28-pixel visual patch (rounded up per side). Example: 500×500 ≈ 324 tokens.
• Gemini (2.x / 3 family): 258 tokens per 768×768 tile (small images count as one tile). Example: 500×500 ≈ 258 tokens.

These are input-side estimates before wrapper instructions, prompt caching, provider downscaling of oversized images, output cost, latency and retries. Older model families (for example tile-based GPT-4o-era billing) calculate the same image differently.

The risk score is deterministic and runs entirely in your browser. It scans for exact identifiers (UUIDs, long hex hashes, URLs, email addresses, file paths, precise decimals, version numbers, phone numbers), credential-like strings, and code-heavy content, and combines them with the fitted font size of the rendered card:

• Low: prose and contextual material where gist matters.
• Medium: some exact identifiers or code — keep those parts as text.
• High: credential-like strings or many exact identifiers — do not compress.

Prompt contents are never uploaded or logged; the analysis is local.

The recommended layout compares square page sizes (512, 768, 1024) using a conservative characters-per-page capacity at a model-readable font, and picks the page size and count with the lowest estimated image tokens. It intentionally refuses to recommend microtext: if a prompt cannot fit legibly, we recommend more pages or plain text.

• No universal savings percentage. Savings depend on model, prompt length, layout and task. Published research reports reductions ranging from roughly 36% to 91% with significant variance — which is why we show a per-prompt, per-model estimate instead of one number.
• No accuracy guarantee. We have not yet published task-success benchmarks. Until we do, verify that your model reads your images correctly before relying on them. Token reduction without task success is not savings.
• No subscription-bill savings. Flat-rate ChatGPT/Claude plans are not billed per token; direct dollar savings apply to metered API usage.

What is visual prompt compression? · Image token calculator · When not to use images