Computer vision
The AI field focused on letting machines understand images and video — covers object detection, image classification, segmentation, OCR, scene understanding, and more.
Computer vision predates LLMs by decades, but the 2024–2026 shift has been dramatic: multimodal LLMs (GPT-4o, Claude Sonnet 4.6, Gemini 2.5) now handle most computer vision tasks via prompt, replacing what used to require specialized models and pipelines.
For specialized domains (medical imaging, satellite analysis, industrial defect detection), purpose-built vision models still win on accuracy and cost-per-inference. For general-purpose vision tasks (reading a screenshot, describing a scene, extracting structured data from an image), multimodal LLMs are now the default.
For agent builders, computer vision means giving the agent eyes. Browser agents read screenshots to decide what to click; document agents read PDFs and forms; multimodal assistants read photos. The capability is foundational for any agent operating in visual environments.
Frequently asked
Do I need a specialized CV model in 2026?+
For general image understanding: no, use a multimodal LLM. For specialized domains (medical imaging, defect detection, security camera analysis): yes, purpose-built models still lead. For high-volume cost-sensitive workloads: smaller specialized models beat multimodal LLMs on cost.
Can computer vision handle video?+
Modern multimodal LLMs handle short clips (under 1 minute) directly. Longer video requires sampling key frames or specialized video models. Real-time video reasoning is an active research area in 2026.