Convert text into images. Cut Claude bills 60% on dense inputs.

Anthropic charges images by pixel dimensions, not characters. Dense reflowed text packed into a 1568×1568 PNG costs roughly the same as ~3,000 text tokens — regardless of whether it holds 3,000 or 15,000 characters. On long code, docs, transcripts, and logs, this is a real cost cut.

Trade-off: the model reads back via OCR, which is lossy. Exact strings (hashes, UUIDs, API keys, hex, exact numbers) are auto-detected and kept as text so nothing critical gets mis-read. Don't use this for tasks requiring byte-exact recall of the packed content.

Input text

density12px

width1440

Rendered image

— × — — image tokens 1 page

Cost comparison

as text

—

— tokens

as image

—

— tokens

saved

—

— %

Exact-recall spans (kept as text, auto-detected)

Paste input to see what stays text and what becomes image.

ready · paste text + prompt

History

Every send is logged locally with cost and savings. Nothing leaves your device.

Settings

Anthropic API key

Stays in your browser's IndexedDB. Never sent anywhere except direct-to-Anthropic. You can wipe it any time.

Prices ($/1M tokens · July 2026)

Editable. Images bill at the model's per-token input rate — image tokens ≈ ⌈W/28⌉ × ⌈H/28⌉.

Haiku

Sonnet

Opus

Auto-preserve patterns

Regex patterns that stay as text (won't be image-rendered). Comma-separated.

Clear history

Wipes every logged send. Irreversible.

Wipe everything

Removes API key, history, settings.

How this works

The technique. Anthropic bills images at ⌈width/28⌉ × ⌈height/28⌉ visual tokens. The token count is fixed by pixel dimensions — not by how many characters are inside the image. Pack dense reflowed monospace text into a 1568×1568 PNG and you get ~3,000 tokens for what would be ~15,000 text tokens.

The research. DeepSeek-OCR (Oct 2025) showed vision models can decode text from a small set of visual tokens at ~10× compression while holding ~97% OCR precision. The pxpipe tool (Dec 2025) demonstrated the practical Claude application.

Where it shines. Dense code review, log analysis, doc summarisation, transcript ingestion, long-context Q&A where the model just needs a rough read of the content.

Where it breaks silently. Byte-exact recall. Model misreads hashes, UUIDs, hex, exact numbers. Reported OCR failure rates: 13/15 correct on Fable, 0/15 on Opus for exact hex recall. FallPx auto-detects these patterns and keeps them as text — but the technique fundamentally isn't suited for tasks needing pixel-perfect recall of the packed content.

Window, not floor. Anthropic can reprice image tokens or add anti-abuse limits at any time. Treat the savings as a temporary edge, not a permanent baseline.

Sovereignty. Your API key stays in this browser's IndexedDB. Requests go direct-to-Anthropic — nothing routes through us. Wipe history any time. MIT.