Code, JSON, and system logs burn through tokens quickly. Stop paying text rates. Encode your data into dense pixels and let Claude's vision encoder do the heavy lifting.
Start Optimizing for Free| Data Size | Standard Text Tokens (Cost) | PXINK Image Tiles (Cost) | Savings |
|---|---|---|---|
| 25,000 chars of Code | ~6,250 tokens ($0.018) | 1 Image (approx $0.002) | 88% |
| 100,000 chars of JSON | ~30,000 tokens ($0.090) | 4 Images (approx $0.008) | 91% |
| 1,000,000 chars of Logs | ~350,000 tokens ($1.050) | 40 Images (approx $0.080) | 92% |
*Costs estimated based on Claude Sonnet 4.6 input pricing at $3.00 per million tokens. JSON and code typically have terrible token-to-character ratios (1 token per 3-4 chars). Images are priced by tile, regardless of how dense the text inside them is.
Yes. You are simply utilizing their multimodal vision capabilities to read data. Anthropic encourages efficient use of their APIs, and vision encoders are specifically designed to read text in images.
Yes, but the savings are smaller because Haiku's base text token price is already extremely low. The biggest ROI comes when using powerful models like Claude Sonnet 4.6 or Opus.