Google has released Gemini 3.8 Flash, which the company claims delivers its highest performance to date in inference and coding. The model consists of two variations: a standard version suited for agentic tasks and software development, and "Gemini 3.8 Flash Cyber," which is specialized for vulnerability detection and mitigation.
Model ReleasesPricing & LimitsGoogleGemini 3.8 FlashGemini 3.8 Flash Cyber
Google Announces Gemini 3.8 Flash, Improving Coding and inference Performance
This article is a translation. Read the Japanese original
Gemini 3.8 Flash Cyber replaces the previous 3.5 version. According to Google's internal testing, this model shows significant improvements over previous models in identifying vulnerabilities and creating patches. The Chrome security team reported that the new model improved patch accuracy by 2.6 times.
Regarding API pricing for developers, Google has set introductory pricing until the end of the year at $0.75 per 1 million input tokens and $3.75 per 1 million output tokens. The regular price is expected to be $1.50 for input and $7.50 for output.
In benchmark tests, Gemini 3.8 Flash took the top spot on the DeepSWE leaderboard, which measures the ability to solve software engineering problems. Meanwhile, in the OSWorld-2.0 test measuring computer operation capabilities, it showed improvement over Gemini 3.7 Flash but still fell short of competing models such as Claude Opus.
Gemini 3.8 Flash is available across the Google ecosystem starting today. A Pro or Ultra subscription is required to use it in the Gemini app, though it can be trialed for free via AI Studio. The Cyber model is currently limited to trusted testers and government agencies.
Source: Google releases Gemini 3.8 Flash, its third Flash model in six weeks (Ars Technica AI, 2026-09-03)