Google launches two Gemini 3.8 models with cutting-edge reasoning capabilities

Three weeks after releasing its latest large language model, Google LLC unveiled Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on the same day. The two language models share a technical foundation, with the main difference being Gemini 3.8 Flash's general-purpose design and Gemini 3.8 Flash Cyber's focus on cybersecurity research. The latter model is accessible through the Fairwind Program, a new early access initiative launched simultaneously with the other two LLMs.

Google put Gemini 3.8 Flash to the test, running it through 16 AI benchmarks. The model outperformed Claude Opus 5 and GPT 5.6 Sol in nine of the tests. One notable achievement was in Terminal-Bench 1.1, which assesses LLMs' capability to handle multi-step coding tasks. Gemini 3.8 Flash excelled in this area, accurately completing the task more effectively than its competitors.

The model also performed well on various other benchmarks, including financial data analysis and chart comprehension. On DeepSWE-1.1, which measures AI models' ability to automate long-horizon coding tasks, Gemini 3.8 Flash scored 73.7%, narrowly beating GPT-5.6 Sol (73.6%) and just a fraction behind Claude Opus 5 (73.8%).

Gemini 3.8 Flash Cyber, designed specifically for cybersecurity researchers, also demonstrated competitive performance. It achieved an 86.2% score on the CyberGym benchmark, which evaluates LLMs' ability to identify vulnerabilities in C and C++ code. This result surpassed Claude Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%). The Fairwind Program, through which Gemini 3.8 Flash Cyber is made available, currently has over 650 participants, including major software companies like Snowflake Inc., CrowdStrike Holdings Inc., and Datadog Inc. These early adopters can utilize Gemini 3.8 Flash Cyber with CodeMender, a technology developed by Google DeepMind to enhance the LLM's vulnerability detection, evaluation, and patch development capabilities.

Google internally employs Gemini 3.8 Flash Cyber to expedite its own work, with developers of the Chrome browser finding that the LLM generates "2.6 times more correct patches" for browser vulnerabilities compared to competitors. Additionally, a Google team utilized Gemini 3.8 Flash Cyber to uncover a "critical foundational vulnerability" that would have taken months to discover manually.