Multimodal Model

A multimodal model accepts more than just text — typically images, sometimes audio or video — and responds in text (or, increasingly, in any modality). GPT-4o, Claude 3.5/4, Gemini 2.5, and Llama 3.2 Vision all qualify.

Real production uses: reading invoices and receipts, screenshot-based testing, accessibility descriptions, product photo tagging, video moderation, voice agents. The thing to know: multimodal input costs more tokens than you'd guess. A high-resolution image can be 1,000+ tokens. Resize and crop before sending if you can.