# Gemini 3 Flash - Your Daily Workhorse Upgraded Source: https://www.youtube.com/watch?v=KxXp8oiMzSw Recap page: https://rapidrecap.app/video/KxXp8oiMzSw Generated: 2025-12-17T16:40:29.656+00:00 --- ## Quick Overview Gemini 3 Flash is Google's latest model built for speed, delivering performance on par with or better than Gemini 2.5 Pro and Gemini 3 Pro across various benchmarks while maintaining superior cost-efficiency, making it the ideal daily workhorse model for production systems. **Key Points:** - Gemini 3 Flash achieves 78% on the SWE-bench Verified benchmark, outperforming Gemini 2.5 Pro and nearing Gemini 3 Pro's performance. - Flash is significantly cheaper than Gemini 3 Pro, costing $0.50/$3.00 per million input/output tokens compared to $2.00/$12.00 for Pro (for input/output tokens, respectively). - The model demonstrated strong multimodal capabilities, successfully extracting recipes from food images and detailed information from complex documents like tax forms and resumes. - Token efficiency tests show Gemini 3 Flash (1585 tokens) is more efficient than Gemini 2.5 Flash (1650 tokens) and significantly better than Gemini 3 Pro (1788 tokens) for generating output. - The model supports advanced tools like Structured Outputs (using Pydantic schemas), Code Execution, and Function Calling, enabling complex agentic workflows. - For spatial understanding tasks, Gemini 3 Flash accurately generated 3D bounding boxes for objects in multi-view images, demonstrating robust spatial reasoning. ![Screenshot at 00:09: The official title card for the presentation, 'Gemini 3 Flash', is displayed, signaling the focus of the video on this new, speed-optimized model release.](https://ss.rapidrecap.app/screens/KxXp8oiMzSw/00-00-09.png) **Context:** This video introduces Gemini 3 Flash, Google's new large language model focused on speed and cost-efficiency, positioning it as an upgrade to the Gemini 2.5 series and a lighter-weight alternative to Gemini 3 Pro. The presenter explores its capabilities through demonstrations in coding, multimodal tasks (image analysis and document parsing), and agentic workflows, comparing its performance and cost directly against its siblings and competitors like Claude Sonnet 4.5 and Grok 4.1. ## Detailed Analysis The video announces the release of Gemini 3 Flash, emphasizing its speed and cost benefits while maintaining high intelligence. Benchmarks show Gemini 3 Flash achieving a 78% score on SWE-bench Verified, outperforming Gemini 2.5 Pro and rivaling Gemini 3 Pro, especially when considering cost savings. For instance, Flash costs $0.50/$3.00 per million input/output tokens, drastically cheaper than Pro ($2.00/$12.00). The presenter demonstrates Flash's strong multimodal capabilities, successfully using it to extract structured recipes from an image of Tom Yum soup and parse detailed handwritten fields (like mailing address, ZIP code, and financial data) from a Form 5500-EZ tax document. Furthermore, the model shows strong spatial reasoning by accurately generating 3D bounding boxes for objects in multi-view kitchen images. The demonstration also highlights the model's ability to handle structured outputs via Pydantic schemas for meeting analysis and its capability to perform complex agentic tasks like code execution, suggesting it will become the primary daily workhorse model for many developers. ### Gemini 3 Flash Overview - Release of Gemini 3 Flash, built for speed and cost-efficiency - Performance nearing Gemini 3 Pro while significantly cheaper than 2.5 Pro - Flash is positioned as the daily workhorse model. ### Performance Benchmarks - Achieves 37.5% on Humanity's Last Exam (with tools), outperforming 2.5 Flash (11.0%) and 2.5 Pro (21.6%) - Scores 81.2% on MMLU-Pro, comparable to Gemini 3 Pro's 81.0%. ### Cost Comparison - Input price of $0.50/M tokens vs. Pro at $2.00/M tokens - Output price of $3.00/M tokens vs. Pro at $12.00/M tokens. ### Multimodal Capabilities (Recipe Extraction) - Successfully analyzed an image of Tom Yum soup to output structured recipe data including ingredients, steps, cuisine (Thai), and calorie estimates per ingredient. ### Multimodal Capabilities (Document Parsing) - Successfully parsed handwritten data from a Form 5500-EZ tax document, extracting plan name, mailing address, ZIP code, form number, dates, and financial liabilities. ### Tool Usage & Agentic Workflows - Demonstrates Structured Outputs using Pydantic for meeting analysis, Code Execution, and Function Calling capabilities. ### Spatial Understanding - Successfully identified objects (curtains, toaster, blender, microwave, sink) in 2D images and generated 3D bounding box coordinates for them. ![Screenshot at 00:14: Title card displaying 'Gemini 3 Flash' alongside the Gemini logo, introducing the new speed-optimized model.](https://ss.rapidrecap.app/screens/KxXp8oiMzSw/00-00-14.png) ![Screenshot at 00:50: Comparison charts showing Gemini 3 Flash outperforming competitors like Claude Sonnet 4.5 and GPT-4 on Humanity's Last Exam, SWE-bench, and MMLU-Pro benchmarks.](https://ss.rapidrecap.app/screens/KxXp8oiMzSw/00-00-50.png) ![Screenshot at 00:32: Code snippet defining Pydantic models for structured output extraction from meeting transcripts, showing the capability for complex data structuring.](https://ss.rapidrecap.app/screens/KxXp8oiMzSw/00-00-32.png) ![Screenshot at 04:03: A demonstration of the model generating code to add a contact form to a website using Google Antigravity IDE.](https://ss.rapidrecap.app/screens/KxXp8oiMzSw/00-04-03.png) ![Screenshot at 14:16: Python code demonstrating the use of the Interactions API to perform spatial understanding by requesting 2D/3D bounding boxes for objects in an image.](https://ss.rapidrecap.app/screens/KxXp8oiMzSw/00-14-16.png)