When Multimodal Hurts

Every modality you add costs tokens, latency, and money — before it adds capability. A 1024px image ≈ 1,500+ tokens in modern encoders; a 30-second audio clip is thousands more; video multiplies it by duration.

The bill compounds:

• Context — visual tokens displace text context you might have spent on instructions or RAG results
• Latency — encode-then-reason serializes; time-to-answer grows
• Cost per call — a multimodal query can run 5-50× a text call
• Quality risk — more input surface = more ways to be subtly wrong

Multimodal should answer a question text cannot — otherwise it's a tax with no revenue.