Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on Sept. 2, 2026, calling them its third Flash release in six weeks. Under Google's Frontier Safety Framework, ordinary 3.8 Flash keeps tighter safeguards against offensive uses in cyber and in chemical, biological, radiological and nuclear (CBRN) areas. The Cyber version relaxes its cyber limits and goes only to trusted defenders, so the most permissive version stays out of general hands.
Google calls 3.8 Flash its most intelligent Flash workhorse for long software-engineering jobs and for work done by AI agents, AI that can carry out tasks on their own. Google says it often takes more steps and uses more tools on hard problems. It accepts text, images, video, audio and PDF files. Google AI Pro and Ultra subscribers get it in the Gemini app, AI Mode in Search and Gemini in Sheets. Developers can build with it in Google AI Studio, Android Studio and Google Antigravity, or through the Gemini API, which lets developers plug it into their own apps. Google says 3.7 Flash remains fully supported for work where efficiency comes first.
Developers are billed for how much the model reads and writes, counted in small pieces that are often parts of words. Through Dec. 31, 2026, Google is keeping 3.7 Flash's introductory price: $0.75 per million pieces read and $3.75 per million written. After that, the prices rise to $1.50 and $7.50. Google's report on the model says its knowledge runs to March 2026 in some areas but may stop at January 2025 in others.
Gemini 3.8 Flash Cyber is built on the same model, Google says, but with looser cyber restrictions. Only trusted defenders, such as governments, critical infrastructure operators and software maintainers, can use it, through Google's Fairwind Program. Google says the model finds vulnerabilities, weaknesses in software that an attacker could use, about as well as the most advanced AI models on a standard test, though its launch post gives no number for that. On Google's own test of finding vulnerabilities in code written in 20 programming languages, the company says the model succeeded more than 70% of the time. Google's Chrome security team found it produced 2.6 times as many correct fixes for Chrome flaws as the best commercial models, which are much larger. Google says it put fixing flaws ahead of exploiting them.
Artificial Analysis, an outside firm that scores AI models, gives 3.8 Flash on its high setting a score of 41 on its Intelligence Index, well above the median of 24 for comparable models, and notes that the model writes unusually long answers. According to a news report, Google's own charts show bigger gains in coding than in other areas, and on a test of operating a computer the model still trails leaders such as Claude Opus.
Outside testers have yet to show whether the Cyber model's gains hold up in their hands, and Google has not said how far it will widen Fairwind beyond its first trusted defenders.