.svg){kind=spacer}
Details
- Google AI Developers showcased Gemini 3 Flash handling complex function calling in a demo, sequencing tasks to prepare ramen using reasoning across 100 ingredients and 100 tools simultaneously.
- The demo is accessible in Google AI Studio, highlighting the model's agentic capabilities for intricate workflows.
- Gemini 3 Flash, recently released on December 17, 2025, is now the default model in the Gemini app, offering PhD-level reasoning, multimodal understanding (text, images, audio, video), and Flash-level speed.
- It outperforms Gemini 2.5 Pro on benchmarks like GPQA Diamond (90.4%), MMMU Pro (81.2%), and SWE-bench Verified (78%), while being 3x faster and cheaper at $0.50/1M input tokens.
- Key features include adjustable thinking levels for complex tasks, 30% fewer tokens on average, and strong performance in coding, video analysis, and real-time applications like in-game assistants.
- Adopted by companies like JetBrains, Figma, Salesforce, and Box for efficient reasoning in high-volume processes.
Impact
Google's Gemini 3 Flash demo underscores its push toward agentic AI that rivals larger models like Gemini 3 Pro in reasoning and tool use while maintaining low latency and cost, pressuring competitors such as OpenAI's GPT-4o mini and Anthropic's Claude 3.5 Haiku, which lag in benchmarks like SWE-bench where it scores 78%. This efficiency lowers barriers for developers building production agents, potentially accelerating adoption in real-time apps from customer support to gaming, and shifting market dynamics by enabling high-frequency workflows without GPU bottlenecks. As it rolls out globally in Search's AI Mode and Vertex AI, it aligns with trends in on-device inference and multimodal agents, likely steering R&D toward scalable intelligence over raw scale in the next 12-24 months, with enterprises like Figma already integrating it to streamline design-to-code processes.
