Google launches Gemini 3.8 Flash for coding agents and cybersecurity

Google DeepMind announced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on September 2, 2026, positioning the pair as faster, lower-cost models for agentic software work and cybersecurity. The Google DeepMind Blog says this is the company’s third Flash release in six weeks.

Gemini 3.8 Flash is the general-purpose option. Google describes it as its strongest reasoning and coding model yet, with improvements in software engineering, multi-step analysis and autonomous task execution. Pricing remains the same as Gemini 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens.

The model is designed to spend additional compute on difficult jobs, taking more reasoning steps and making repeated tool calls. That could help developers build agents that handle longer workflows, such as completing an engineering task from start to finish. Google says 3.8 Flash outperformed most larger frontier models on the DeepSWE v1.1 long-horizon software engineering evaluation. It also reports a 54.9% score on HLE-Verified, alongside gains on finance and legal agent benchmarks.

That extra diligence may increase token use, so developers can lower the model’s effort setting when efficiency matters—or continue using 3.7 Flash for cost-sensitive workloads. The unchanged introductory price makes 3.8 Flash especially notable for teams experimenting with autonomous coding systems.

Gemini 3.8 Flash Cyber is aimed specifically at vulnerability discovery and automated patching. Google calls it a frontier-level cybersecurity model, but access is limited to trusted defenders through the Fairwind Program. Both variants share the same underlying model and were improved through training and agentic evaluation loops, including demanding cybersecurity work.

Source: Google DeepMind Blog