Google DeepMind Releases Gemini 3.8 Flash and Gemini 3.8 Flash Cyber: One Core Model, Two Access Envelopes

Google DeepMind released two variants of the same foundational AI model that differ only in their safety restrictions rather than their underlying architecture. The standard version is available through multiple platforms for general use, while the specialized cybersecurity version is restricted to vetted defenders such as government authorities and infrastructure operators through a gated access program. The new model executes additional reasoning steps and tool calls on complex tasks, which improves accuracy but increases token consumption compared to its predecessor. Performance improvements are demonstrated through benchmarks in software engineering, vulnerability discovery, and patch generation, with the cybersecurity variant positioned as cost-competitive with larger models on defensive security tasks.