Why are AI vision tokens so much cheaper than text tokens for long context windows? 💸

Anthropic and Google charge heavy text token rates when you feed them huge context windows, logs, or codebases. But if you render that exact same text into a dense PNG, vision OCR pricing kicks in—cutting overall token costs by ~60%.

Is vision pricing just massively subsidized right now by AI providers, or are text tokenizers fundamentally overpriced for long, dense payloads?

What other weird billing quirks or API pricing anomalies have you run into while building?