Introducing Gemini 3.8 Flash and 3.8 Flash Cyber

Details

Gemini 3.8 Flash is Google's third Flash model in six weeks. The progress from 3.7 is concrete: 54.9% on HLE-Verified (multi-step STEM, humanities, and professional reasoning), top performance on DeepSWE v1.1 long-horizon software engineering (beating most larger frontier models), and leading scores on Vals Finance Agent V2 and Harvey's Legal Agent Benchmark — all at identical pricing to its predecessor.\n\nThe model is designed to 'work harder' on complex tasks by executing additional reasoning steps and tool calls iteratively. This means token consumption can increase on difficult problems, particularly at higher effort levels. Developers can dial down effort levels to minimize token overhead where throughput matters more than peak quality. Gemini 3.7 Flash remains available for compute-constrained workloads.\n\nThe 3.8 Flash Cyber variant, powering the Fairwind Program for cyber defense, shares the same foundational training but is specialized for vulnerability detection and autonomous patching. The rapid three-release-in-six-weeks cadence signals sustained iteration pressure that competing labs will need to match.