The model is designed for longer-running software engineering and autonomous agent workflows. On DeepSWE v1.1, which evaluates end-to-end software engineering tasks, Google says Gemini 3.8 Flash outperformed most larger frontier models while operating at a lower cost.
Google also reported gains across finance, legal and broader reasoning benchmarks. Gemini 3.8 Flash improved on its predecessor in Vals Finance Agent V2 and Harvey’s Legal Agent Benchmark, while scoring 54.9% on HLE-Verified, a test covering multi-step reasoning across professional, scientific and humanities domains. Those gains come with a tradeoff. Google says Gemini 3.8 Flash may use more reasoning steps and make repeated tool calls on difficult tasks, particularly at higher effort settings. Developers focused more heavily on efficiency can reduce effort levels or continue using Gemini 3.7 Flash.
Google demonstrated the model building several interactive applications, including a 3D game, a DOS-style version of Google Maps, a topographic visualization using U.S. Geological Survey data and a hardware teardown visualizer created in Google AI Studio.
Gemini 3.8 Flash Cyber uses the same core intelligence but is tuned specifically for defensive cybersecurity work. Google says the model achieved frontier-level results on CyberGym, a vulnerability discovery benchmark, and exceeded a 70% success rate on an internal test spanning complex codebases across 20 programming languages.
Patching is another major focus. On CWE-Bench, Gemini 3.8 Flash Cyber reached a pass@1 score of 47.2%, compared with 47.8% for a leading frontier model, while Google says it operates at substantially lower cost.
The company is already using the cyber model internally. Google says its Chrome Security team saw 2.6 times more correct vulnerability patches than with larger commercial models. Its Cloud Vulnerability Research team also used Gemini 3.8 Flash Cyber to identify a critical foundational vulnerability in less than two hours.
Wiz reported that the model achieved between 7.5% and 9.7% higher recall on an internal penetration-testing benchmark while costing between 2.3 and 5.2 times less than other leading frontier models.
Access to Gemini 3.8 Flash Cyber is more restricted because it has fewer cybersecurity limitations than the standard model. Google is making it available through its new Fairwind Program to selected government authorities, critical infrastructure operators and software maintainers.
Both models include safeguards covering chemical, biological, radiological and nuclear risks as well as cyber misuse under Google’s Frontier Safety Framework. Google also says the Gemini 3.8 models improved substantially in resistance to prompt-injection attacks based on testing by Gray Swan.
Gemini 3.8 Flash is available to developers through the Gemini API in Google AI Studio and Android Studio, as well as through Google Antigravity and Stitch. Enterprises can access it through Gemini Enterprise, while Google AI Pro and Ultra subscribers can use the model in the Gemini app, AI Mode in Google Search and Gemini in Google Sheets.
This analysis is based on reporting from Google.
Image courtesy of Google.
This article was generated with AI assistance and reviewed for accuracy and quality.
