Stop thinking of Vision APIs as just for "identifying objects in photos". Vision is a high-bandwidth, low-cost data pipe straight into the LLM's brain.
Start Optimizing for Free
When you send text to an LLM via the standard `content` array, it goes through a tokenizer (like BPE or Tiktoken). For code, JSON, and UUIDs, this tokenization is brutally inefficient.
When you send an image, the LLM processes it in "patches" (usually 14x14 or 28x28 pixels) using a Vision Transformer (ViT). The billing model charges you per patch, not per word. By rendering deterministic, non-anti-aliased 5x8 pixel fonts into those patches, we can pack roughly 6 characters into a single token's worth of image space.
pxpipe is the open-source rendering engine behind PXINK. It bypasses the browser's standard text rendering (which uses sub-pixel anti-aliasing) in favor of a raw bitmap atlas to ensure 100% crisp pixels for AI OCR.
Yes. You can use the `canvas` npm package to run the exact same rendering logic in a Node.js backend environment.