Overview
- Google quietly posted Gemini 3.8 Flash on its DeepMind site on Wednesday and announced a Cyber‑specialized 3.8 Flash variant for security teams.
- The company says internal benchmarks show 3.8 Flash meaningfully improves coding performance versus prior Gemini releases and that engineers preferred it over a rival model in internal Jetski tests.
- Gemini 3.8 Flash supports very large context windows up to 1 million tokens and up to 64K text output, and offers tunable inference effort so users can trade cost, latency and quality.
- Google reports the 3.8 Flash Cyber model has been deployed internally for code‑security defense, achieved over 70% vulnerability‑finding success in company tests and will be offered to trusted teams through the Fairwind program.
- Separately, Google rolled out an agentic video understanding feature across Flash models via Google AI Studio and the Gemini API to cut token use and cost in long‑video tasks and plans to extend it to YouTube; the models still carry a March 2026 knowledge cutoff and retain common LLM limits such as occasional hallucinations.