Gemini Flash 3 is my new favorite model (yes really)

Gemini Flash 3 outperforms its predecessor 2.5 Flash and rivals much more expensive models like Claude Opus 4.5 in benchmarks, particularly excelling at spatial reasoning, multimodal processing (images, video, audio, PDFs), and bulk data analysis. While significantly faster and cheaper than Gemini 3 Pro, it generates 2x more reasoning tokens than 2.5 Flash, making it more expensive than expected. The model has a 91% hallucination rate when uncertain and struggles with instruction-following compared to OpenAI and Anthropic models. Best suited for structured data processing tasks with JSON outputs rather than chat or coding applications.