Google launches Gemini 3.8 Flash and Gemini 3.8 Flash Cyber

Google introduced two AI models on Wednesday : Gemini 3.8 Flash, a general-purpose model aimed at software engineering and autonomous tasks, and Gemini 3.8 Flash Cyber, a variant built for cybersecurity work. Google noted that the launch comes three weeks after Gemini 3.7 Flash and represents the third Flash-generation release this quarter.

Gemini 3.8 Flash is priced at $0.75 per million input tokens and $3.75 per million output tokens — the same introductory rate as its predecessor — and is available to Google AI Pro and Ultra subscribers through the Gemini app, in AI Mode in Google Search, in Gemini in Google Sheets, and to developers via the Gemini API and AI Studio, the company said. On the DeepSWE v1.1 long-horizon software engineering benchmark, the model outperforms most larger frontier models while costing less, according to Google. It also achieves a 54.9% score on HLE-Verified, a benchmark spanning reasoning across STEM, humanities, and professional fields.

Google attributed the improvements to an architectural decision that pushes the model to work harder on complex problems — iterating through reasoning chains and invoking tools multiple times — which can drive up token consumption when the effort setting is high. Developers prioritizing efficiency can use lower effort settings or continue using Gemini 3.7 Flash, the company added.

Gemini 3.8 Flash Cyber is available only to participants in Google's new Fairwind Program, which the company said will provide access to government authorities, critical infrastructure operators, and software maintainers. On CyberGym, an industry benchmark for autonomous vulnerability discovery, the Cyber model surpasses both its predecessor and larger frontier models, Google said. On CWE-Bench, an external benchmark for patching, it achieves a pass@1 rate of 47.2%, compared with 47.8% for a leading frontier model, at lower cost.

Google said it is already deploying the Cyber model internally. According to Google, the Chrome Security team's testing showed the Cyber model generated correct vulnerability patches at a rate 2.6 times higher than competing commercial models of greater size. Google's Cloud Vulnerability Research team also reported that the model surfaced a critical foundational vulnerability in less than two hours — work that ordinarily requires months of research.

The release follows reporting that Google has been testing the model internally against Anthropic's Opus model using Jetski, Google's internal coding tool, with engineers preferring the new model. Google has devoted more resources to reinforcement learning since the start of the year, a training approach that teaches models to perform skills through trial and error. Koray Kavukcuoglu, who became chief executive of Google DeepMind last month after co-founder Demis Hassabis stepped aside, has signaled he wants to increase the pace of execution.