Gemini 3.8 - Google's Reasoning and Coding Model

What is Gemini 3.8

Gemini 3.8 is a reasoning and coding model launched by Google, comprising two versions: Flash and Flash Cyber. Gemini 3.8 Flash, while maintaining the same speed and low price as its predecessor, significantly improves software engineering, agent tasks, and multi-step reasoning capabilities, approaching or even surpassing larger-parameter frontier models on benchmarks in code, finance, law, and other domains. Gemini 3.8 Flash Cyber is a version built specifically for cybersecurity, with top-tier vulnerability discovery and automated patch remediation capabilities, available only to certified security defenders through the Fairwind Program.

Key Features of Gemini 3.8

• Long-horizon software engineering: Autonomously completes end-to-end solutions to complex engineering problems, performing close to or even surpassing larger-parameter frontier models on benchmarks such as DeepSWE.
• Autonomous agent tasks: Supports complex multi-step agent workflows, capable of iterative tool calling and reasoning, suitable for scenarios such as terminal coding and computer operation.
• Domain-specific reasoning: Demonstrates enterprise-grade reliability in financial analysis, legal workflows, and interdisciplinary expert reasoning.
• Cybersecurity offense and defense: The Flash Cyber version focuses on autonomous vulnerability discovery and automated patch remediation, reaching frontier levels on industry benchmarks such as CyberGym and CWE-Bench.
• Native multimodality: Supports multiple input modalities such as text, code, images, and video, and can generate rich media content such as 3D visualizations, games, and interactive applications.

Technical Principles of Gemini 3.8

• Long-running agentic loops: Gemini 3.8 introduces long-running agentic loops, a mechanism that enables the model to recursively evaluate and refine its own output during task execution. Unlike a single forward pass, the model continuously iterates on complex tasks—performing additional reasoning steps, repeatedly calling external tools, and dynamically adjusting strategies based on intermediate results—achieving deeper execution and higher accuracy in multi-step reasoning and code generation tasks.
• Dynamic compute investment and reasoning effort adjustment: The model adopts the design philosophy of "investing more compute," demonstrating greater diligence on complex tasks. The model automatically allocates more reasoning resources based on task difficulty, and at higher effort levels may consume more tokens to maximize performance; at the same time, developers can choose lower reasoning effort levels to reduce token costs, flexibly balancing performance and cost.
• Deep specialized training in cybersecurity: Gemini 3.8 Flash and Flash Cyber share the same underlying base intelligence, and the significant improvements in coding and reasoning capabilities stem in part from rigorous training in the highly demanding domain of cybersecurity. Through deep optimization on tasks such as vulnerability discovery and patch remediation, the model has learned to identify and fix complex code vulnerabilities across 20 programming languages, internalizing a defender-first security mindset.

How to Use Gemini 3.8

• Developers: Call 3.8 Flash through Google AI Studio, the Gemini API, or Android Studio, or explore agent-first workflows in Google Antigravity , or generate UI in Stitch.
• Enterprise users: Directly integrate 3.8 Flash in the Gemini Enterprise platform to build enterprise-grade intelligent applications.
• Ordinary consumers: After subscribing to Google AI Pro or Ultra, you can use 3.8 Flash in the Gemini App, AI Mode in Google Search, and Google Sheets.
• Security defenders: Submit an application through the Fairwind Program. After certification, you can get priority access to 3.8 Flash Cyber for vulnerability discovery and automated patch remediation.

Core advantages of Gemini 3.8

• Extreme cost-effectiveness: It provides reasoning and coding capabilities close to or even surpassing larger-parameter frontier models such as Claude Opus 5 and GPT-5.6 at the low price of the Flash series.
• Long-horizon software engineering: In long-horizon software engineering benchmarks such as DeepSWE v1.1, it can autonomously solve complex engineering problems end to end, with performance ranking among the top in the industry.
• Enterprise-grade professional reasoning: It demonstrates reliability in supporting autonomous tasks in key areas such as financial analysis, legal workflows, and interdisciplinary expert reasoning.
• Dynamic compute adjustment: It supports flexibly adjusting reasoning intensity according to task complexity, allowing more compute resources to be invested in highly difficult tasks to maximize performance, while also reducing Token consumption in efficiency-first scenarios.
• Cybersecurity specialization: The Flash Cyber version reaches the frontier level in vulnerability discovery and automated patch remediation, at a cost far lower than similar large-parameter models.
• Native multimodal and Agent capabilities: Based on a long-running Agent loop mechanism, it supports multimodal inputs such as text, code, images, and video, and can autonomously iterate to complete complex multi-step tasks.

Project address of Gemini 3.8

• Project official website：https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/

Comparison of Gemini 3.8 with similar competitors

Application scenarios of Gemini 3.8

• Autonomous software development: The model supports end-to-end completion of complex code refactoring, Bug fixes, and feature development, and supports long-horizon software engineering tasks.
• Financial quantitative analysis: Automatically parses financial reports and market data, and generates investment research reports, risk assessments, and compliance recommendations.
• Legal document review: Quickly analyzes contract terms, precedents, and regulations to assist lawyers in completing due diligence and compliance review.
• Cybersecurity defense: Automatically scans cross-language codebases for vulnerabilities and generates fix patches, helping security teams respond quickly to threats.
• Multimodal intelligent applications: Generates 3D interactive games, data visualization applications, or rich media content based on natural language.