Gemini 3.5 Flash

Model Overview

Gemini 3.5 Flash is a high-efficiency multimodal model from Google DeepMind, released on May 19, 2026. Designed specifically for agentic workflows, complex reasoning, and coding tasks, it is a frontier model that balances high-speed performance with capabilities that rival or exceed larger flagship models. As part of the Gemini 3 family, it serves as the default "workhorse" model for the Gemini app and AI Mode in Google Search, delivering "near-Pro" intelligence at a "Flash-tier" cost and speed.

Capabilities

• Agentic & Coding Performance: Built for autonomous agent loops and multi-step workflows. It excels at planning and executing complex actions across various environments.
• Multimodal Reasoning: Processes text, images, audio, video, and code. It is highly capable in chart-and-data reasoning and complex visual tasks like document understanding and object counting.
• Computer Use: Features native "computer use" capabilities, allowing the model to interact with browser, desktop, and mobile environments to automate long-horizon tasks.
• Massive Context Window: Supports a 1,048,576 token (approx. 1 million) context window, enabling it to ingest and process vast amounts of data, such as entire codebases, long meeting transcripts, or books in a single prompt.

Example Use Cases

• Autonomous AI Agents: Powering multi-step, complex agentic workflows that require continuous planning and tool execution.
• Codebase Analysis & Generation: Analyzing massive code repositories, identifying bugs, and generating production-ready code.
• Data Extraction & Multimodal Search: Extracting structured data from hours of video, audio transcripts, or hundreds of document pages simultaneously.
• Desktop Automation: Interacting directly with graphical user interfaces to automate repetitive workflows on computers or mobile devices.

Performance & Benchmarks

Gemini 3.5 Flash has demonstrated significant leadership in benchmarks focused on agentic behavior and coding:

• Terminal-Bench 2.1: Scored 76.2%, outperforming several previous flagship models.
• MCP Atlas: Achieved 83.6% in multi-step workflows.
• SWE-Bench Pro: Achieved 55.1% in agentic coding tasks.
• CharXiv Reasoning: Scored 84.2%, surpassing Gemini 3.1 Pro (83.2%).
• MMMU-Pro (no tools): Scored 83.6%.
• Vision: Set a record for the highest score ever recorded on the Roboflow Vision Evals leaderboard.

Intended Use & Limitations

• Intended Use: Fast, scalable, and complex multimodal and agentic tasks where high throughput and low latency are critical.
• Limitations: While exceptionally fast and capable, extremely specialized or deeply nuanced creative tasks might still occasionally benefit from a larger "Pro" or "Ultra" tier model depending on the workload.

About Google DeepMind

Google DeepMind is a premier artificial intelligence research laboratory, formed by the merger of Google Brain and DeepMind. Their mission is to solve intelligence to advance science and benefit humanity, creating general-purpose AI systems like the Gemini family that can learn, reason, and interact safely and efficiently with the world.