Multimodal AI unlocks new application categories by combining vision, audio, and language. GPT-4o Vision for image understanding: base64 encode images or use URLs, include in messages content array alongside text. Use cases: product image analysis, document understanding, UI screenshot debugging, chart interpretation, medical image description. Claude vision: handles complex multi-page documents, diagrams, and detailed image analysis. Excellent for technical drawing interpretation. Gemini 1.5 Pro: 1M token context window supports very long videos or entire codebases as images. Architecture patterns: 1) Image captioning for search indexing: batch process product images to generate searchable descriptions. 2) Document understanding: combine OCR text + layout structure from visual analysis for better extraction. 3) Visual QA: answer user questions about uploaded images with context from knowledge base. 4) Video analysis: sample frames at regular intervals, analyze each, synthesize insights. 5) Audio + text: Whisper transcription + LLM analysis pipeline for meeting intelligence. 6) Medical imaging: describe findings from radiological images, flag for physician review. Implementation: use base64 for images (<20MB), URLs for larger or CDN-hosted media. Batch processing for offline analysis. For real-time: stream from webcam or file upload. Cost: vision tokens are 2-3x text token cost - optimize by resizing images (1024x1024 sufficient for most tasks).