- Published on
Google’s Gemini 3.8 Flash is built for agents, while its Cyber twin hunts vulnerabilities
- Authors
- Name
- DeepHorus
Google has announced the launch of its latest AI models, Gemini 3.8 Flash and its cybersecurity-focused counterpart, Flash Cyber. The new release marks the company's third Flash model in just six weeks, following version 3.7, and is designed to significantly advance capabilities in agentic tasks, software development, and multi-step reasoning. According to Google CEO Sundar Pichai, the models represent "significant leaps" in performance while maintaining the efficiency of the Flash series.
The standard Gemini 3.8 Flash model is positioned as a "workhorse" for complex operations. It has demonstrated superior performance on the DeepSWE coding benchmark, outperforming many larger frontier models at a lower cost. The model features a 1-million-token input window and a 64,000-token output limit, capable of processing text, images, audio, video, and PDFs. Google notes that 3.8 Flash "works harder" and exhibits "greater diligence," sometimes using more tokens to execute additional reasoning steps for complex tasks.
Gemini 3.8 Flash is immediately available for developers through the Gemini API in platforms like Google AI Studio and Android Studio. It is priced at $0.75 per million input tokens and $3.75 per million output tokens, matching the introductory rate of its predecessor. Users can adjust effort levels to balance quality, cost, and latency, and can continue using the fully supported 3.7 Flash for efficiency-focused workloads.
On benchmarks, 3.8 Flash achieved notable results across multiple domains. It scored 54.9% on Humanity's Last Exam, a test of multi-step reasoning across math, science, and the humanities. It also topped specialized benchmarks like Harvey's Legal Agent Benchmark and the Vals Finance Agent V2. In the Agent Arena, it ranked No. 14, above DeepSeek-V4-Pro, and debuted at No. 7 in the Text Arena, ahead of Claude Opus 5.
The cybersecurity model, Flash Cyber, is described as Google's "most capable" security AI. It achieved 86.2% on the CyberGym benchmark and 47.2% on CWE-Bench for patching abilities. In an internal test, it discovered vulnerabilities across 20 programming languages with over 70% success. Google highlights its utility for identifying and mitigating security flaws at scale.
Google showcased 3.8 Flash's creative and technical potential through examples, including building a 3D puzzle game from a simple prompt and creating a functional DOS version of Google Maps. These demonstrations underline the model's ability to handle intricate, multi-step projects. The rapid iteration—three Flash models in six weeks—signals Google's aggressive push to lead in the fast-evolving AI agent and cybersecurity space.
