Google launched Gemini 3.8 Flash AI model today. This is the third release in the Flash series in the past six weeks. According to the developers, the new model has become the smartest model in the series and features significant improvements in programming, working with autonomous artificial intelligence agents, and solving complex multi-stage problems.
Meanwhile, Google launched Gemini 3.8 Flash Cyber, a dedicated version of the model for cybersecurity tasks. It’s designed to hunt for vulnerabilities and automatically prepare patches, and is currently only available to trusted experts through the new Fairwind program. Both models are built on the same basic system, but Flash Cyber is less restrictive for tasks related to protecting computer systems.
This update is the third in the past three months. According to Google, this “The smartest Flash model designed for long-term software development, autonomous agents and complex enterprise workflows”.
Google notes that, like previous models, the Gemini 3.8 Flash knowledge deadline is March 2026 “For some areas, while in others [пользователи] Knowledge of the likely discovery model will be limited to January 2025. “
Google says the Gemini 3.8 Flash is a significant improvement over its predecessor, the Gemini 3.7 Flash, and in some tests it’s close to larger, more expensive rival models. In particular, the new version shows good results in long-term software development tasks, when artificial intelligence must independently carry out the entire chain of actions from analyzing the problem to creating and testing the solution.
The model also improves work on specialized tasks that require analyzing large amounts of information and multiple sequential steps. Gemini 3.8 Flash scored 54.9% on the HLE-Verified test, which tests the ability to solve complex problems in the sciences, humanities and professional fields.
At the same time, Google pointed out that the new model is able to spend more computing resources on complex tasks: it performs additional inference steps and repeatedly calls external tools. If speed and efficiency are more important for a specific task, developers can reduce the “workload” rating of the model.
A standalone version of Gemini 3.8 Flash Cyber is targeted at information security professionals. She knows how to independently search for bugs in code and come up with ways to eliminate them. In CyberGym tests, which are used to evaluate AI’s ability to find vulnerabilities, the model performed on par with today’s best systems and beat Flash Cyber’s previous version 3.5, as well as some larger models.
In Google’s own tests, the model found vulnerabilities in complex projects written in 20 programming languages and successfully completed more than 70% of the tasks. In the CWE-Bench test, which evaluates the ability to automatically remediate detected vulnerabilities, Gemini 3.8 Flash Cyber scored 47.2% on the first try, almost on par with the leading model’s 47.8%, but at a significantly lower cost to use.
Google is already using the new model to protect its software. For example, Chrome experts achieved 2.6 times more correct bug fixes with their help than using the best business model, and at a much greater scale. Additionally, the Google Cloud team discovered a critical vulnerability using Gemini 3.8 Flash Cyber in less than two hours—such research typically takes months.
Gemini 3.8 Flash is available to developers and Google AI Pro and Ultra subscribers through Google AI Studio and API. The issuance price is the same as Gemini 3.7 Flash: $0.75 for 1 million input tokens and $3.75 for 1 million output tokens. Gemini 3.8 Flash Cyber is only available to verified cybersecurity professionals through the Fairwind program.
If you find an error, select it with your mouse and press CTRL+ENTER.
