48,000 characters in 2,700 tokens: lets discuss how LLMs read text as images

Short summary

Multimodal LLMs price images by pixel dimensions rather than character count, making dense text dramatically cheaper when rendered as PNG images. pxpipe demonstrated an 89% token reduction ($42 text request → $4 with imaged context), moving vision-based reading from a party trick to viable engineering optimization. Each provider (Anthropic, OpenAI, Google) uses different image token pricing, requiring provider-specific optimization for render sizes and tiling.

• •Image tokens are far more expressive than text tokens—89% token savings possible by converting dense text to PNG
• •Vision encoders accurate enough for production; each provider has different pricing models requiring optimization
• •Practical implications: render size, tiling strategy, and multi-page splitting are now key cost-control levers

Generated with AI, which can make mistakes.