Gemini 3.8 Flash tops DeepSWE, cheaper coding AI

Google's release of Gemini 3.8 Flash marks its third Flash model in six weeks, a rapid cadence that suggests an aggressive push in the AI model race. The model's top position on the DeepSWE leaderboard—which measures software engineering problem-solving—signals that Google's smaller, more efficient Flash models can now compete with premium counterparts, especially after the reported delay of Gemini 3.5 Pro due to coding performance issues. However, the OSWorld-2.0 results reveal that computer use remains a weak spot for Google, with Gemini 3.8 Flash still far behind Claude Opus, though GPT also underperforms in that area.

The Cyber variant, while niche, shows promise in specialized fields. Google claims it identified more vulnerabilities and issued working patches more often than previous models. The Chrome security team saw a 2.6x increase in patch accuracy, and the Cloud team reports that it found a critical vulnerability in just two hours. These claims are backed by partners like Wiz and Palo Alto Networks. Yet, the Cyber model's availability is restricted to trusted testers and governments, indicating that its deployment is carefully controlled for security reasons.