Google launches two Gemini 3.8 models with cutting-edge reasoning capabilities
SiliconANGLE Maria Deutscher ● Covered by 3 sources
Google just shipped Gemini 3.8 Flash and a cybersecurity version built on the same core. The twist: it’s already beating some big rivals on coding and vuln-hunting tests.
Based on reporting by SiliconANGLE, Maria Deutscher — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google has launched two new Gemini models, just three weeks after its last large language model release. Gemini 3.8 Flash is the general-purpose one. Gemini 3.8 Flash Cyber is aimed at cybersecurity researchers, and it arrives alongside a new early access effort called the Fairwind Program.
Google says Gemini 3.8 Flash went through 16 AI benchmarks and came out ahead of Claude Opus 5 and GPT 5.6 Sol on nine of them. The model did particularly well on Terminal-Bench 1.1, a test of multistep coding work, and also held up on benchmarks covering things like financial data analysis and chart reading. On DeepSWE-1.1, which measures long-horizon coding automation, it scored 73.7%.
That number matters because it put Flash about 1% ahead of GPT-5.6 Sol, while still landing a little behind Opus 5. Google’s own explanation for the gains is blunt: 3.8 Flash “works harder.” The company says that means extra reasoning steps and repeated tool use, especially when higher effort is turned on.
Flash Cyber is the more specialized play. Google says it scored 86.2% on CyberGym, a benchmark for finding vulnerabilities in C and C++ code. Claude Mythos 5 scored 83.8%, and GPT-5.6 Sol scored 83.6%. The model is only being offered through Fairwind, which is open to government agencies, critical infrastructure operators and tech firms that help secure widely used software foundations.
Google says more than 650 participants are in the program at launch, including Snowflake, CrowdStrike and Datadog. Those users can pair Flash Cyber with CodeMender, a DeepMind-built harness meant to help models find bugs, judge how serious they are and write patches. Inside Google, the pitch is already turning into a practical one: the Chrome team says the model produced 2.6 times more correct patches than rivals, and another team used it to uncover a critical foundational vulnerability that would have taken months to find by hand.
My take — AI-written commentary, not fact-checked reporting
This is the part where AI stops being a demo and starts being a wrench. Google is clearly betting that “works harder” will sell better than flashier branding, and honestly that’s a healthier pitch than the usual miracle dust. The real tell is the cybersecurity angle: if a model can help patch browsers and dig up vulnerabilities, the market will forgive a lot of benchmark theater.
Read more about this at: SiliconANGLE
