Google Introduces Gemini 3.8 Flash and 3.8 Flash Cyber AI Models
New models deliver significant performance improvements in coding and cybersecurity domains.
1Introduction of Gemini 3.8 Flash
Google has unveiled the Gemini 3.8 Flash and 3.8 Flash Cyber models. These next-gen AI models are optimized for agentic workflows and cybersecurity. They offer significantly improved coding and reasoning capabilities compared to the previous 3.7 Flash version. While maintaining the same speed and low-cost advantage, they approach the performance of higher-cost frontier models.
2Performance Enhancements
In DeepSWE v1.1 tests, 3.8 Flash outperforms most large frontier models in end-to-end complex engineering problem solving. It also delivers impressive results in professional fields like finance and law. The model shows clear superiority over its predecessor in Vals Finance Agent V2 and Harvey's Legal Agent Benchmark tests. It achieves a 54.9% success rate on HLE-Verified tests.
3Design Philosophy
The performance boost of 3.8 Flash stems from a 'work harder' philosophy. It executes additional reasoning steps and uses tools iteratively for complex tasks. While this sometimes leads to higher token usage, it maximizes performance. Developers can choose lower effort levels for compute-efficient applications or continue using 3.7 Flash.
4Creative Applications
3.8 Flash can create games with looping instructions in Google Antigravity. It builds immersive 3D levels using textures generated with Nano Banana. With a single prompt, it can generate a fully functional DOS version of Google Maps. Using USGS data, it creates topographic maps of geographical sites with scientific explanations.
5Hardware Visualization
Developed in Google AI Studio, Hardware Anatomy uses 3.8 Flash to visualize physically-proportioned teardowns of hardware devices in 3D. It features a deconstruction slider that breaks devices into layers. Created with Three.js, these renderings provide a realistic inspection experience.
6Cybersecurity Capabilities
3.8 Flash Cyber is available to trusted defenders through the Fairwind Program. It demonstrates frontier-level performance in autonomous vulnerability discovery on CyberGym tests. In comprehensive internal tests spanning 20 programming languages, it achieves over 70% success rate, significantly outperforming previous models.
7Security-Focused Design
3.8 Flash Cyber equips defenders with expert capabilities to gain advantage over attackers. Vulnerability fixing was prioritized over offensive capabilities during development. On CWE-Bench tests, it offers a cost-effective solution with 47.2% pass@1 rate. Google has already begun using this model to secure its own codebase.
8Pricing and Safety
Gemini 3.8 models show significantly improved prompt injection robustness according to Gray Swan measurements, protecting users from malicious attacks. Introductory pricing is valid until December 31, 2026. Starting January 1, 2027, prices will be $1.50 per 1M input tokens and $7.50 per 1M output tokens. The models incorporate safety measures aligned with the Frontier Safety Framework.