Thursday, September 3, 2026

Google Launches Gemini 3.8 Flash with Enhanced Reasoning and Coding Performance

Input
2026-09-03 09:54:15
Updated
2026-09-03 09:54:15
Google launched Gemini 3.8 Flash, with enhanced reasoning and coding performance, along with Gemini 3.8 Flash Cyber, a model specialized in cybersecurity. Courtesy of Google.

[Financial News] Google has launched Gemini 3.8 Flash with enhanced reasoning, coding, and AI-agent capabilities. It also unveiled Gemini 3.8 Flash Cyber, which is specialized in security operations.
Google said on the 3rd that Gemini 3.8 Flash is its most intelligent flagship model among the Flash-series models introduced to date. The company explained that it had improved the model's performance in software engineering, AI-agent tasks, and multistep reasoning in specialized fields compared with its predecessor, Gemini 3.7 Flash. The successor was released just three weeks after Gemini 3.7 Flash was announced, making it the third Flash model unveiled in the past six weeks.
Gemini 3.8 Flash is designed to perform additional reasoning steps and repeatedly call the tools it needs when handling complex tasks. It focuses on long-running autonomous AI agents and coding work. According to Google, the model outperformed most larger frontier models at a lower cost on DeepSWE v1.1, a benchmark that evaluates long-horizon software engineering capabilities. It also scored 54.9% on the multistep reasoning benchmark HLE-Verified, demonstrating the ability to handle multistep reasoning across STEM, the humanities, and professional practice.
At its launch promotional price, usage costs $0.75 per 1 million input tokens and $3.75 per 1 million output tokens, the same as for its predecessor. Developers can adjust the model's reasoning-effort level according to the nature of a task, balancing performance against token usage.
Gemini 3.8 Flash Cyber, unveiled alongside it, specializes in vulnerability detection and automated patching. In Google's internal evaluation, it achieved a vulnerability-detection success rate of more than 70% on code written in 20 programming languages. On the external patch-generation benchmark CWE-Bench, it recorded a 47.2% patch-generation success rate. Google explained that the model had also been used to identify critical vulnerabilities in core infrastructure within two hours.
Gemini 3.8 Flash is available through Google AI Studio, Android Studio, and Gemini Enterprise. General users can access it through Gemini app and Google AI Mode in Google Search with Google AI Pro or Ultra subscriptions. The cyber model will initially be offered through the Fairwind Program to trusted government agencies, critical-infrastructure operators, and software maintainers.


[email protected] Choi Hye-rim Reporter