Google ships Gemini 3.8 Flash, its third Flash model in six weeks, plus a Cyber variant

机器之心 · wechat · 2026-09-03

Google released Gemini 3.8 Flash, its third Flash-series model in six weeks, priced the same as 3.7 Flash ($0.75/$3.75 per million input/output tokens) with major gains in software engineering, agentic tasks, and multi-step reasoning.

Key points

Performance: Outperforms most larger frontier models on long-horizon SWE benchmark DeepSWE v1.1, scores 54.9% on HLE-Verified, and beats 3.7 Flash on finance and legal agent benchmarks. It spends more compute (more reasoning steps and tool calls); developers can lower effort levels to save tokens.
Gemini 3.8 Flash Cyber: A security-focused variant achieving frontier-level vulnerability discovery on CyberGym, >70% success on an internal 20-language benchmark, and 47.2% Pass@1 on CWE-Bench patching (near frontier's 47.8% at much lower cost). Access via the Fairwind Program.
Real-world use: Chrome's security team saw 2.6x more correct patches vs. larger commercial models; Wiz reported 7.5–9.7% higher recall at 2.3–5.2x lower cost; Google Cloud used it to find a critical vulnerability in under 2 hours that would normally take months.

Available now via AI Studio, Gemini App, and enterprise channels. Google DeepMind's Yao Shunyu called it a big leap for recursive self-improvement.

Original post →

Google ships Gemini 3.8 Flash, its third Flash model in six weeks, plus a Cyber variant

More from Models