Google has unveiled three new artificial intelligence models, including its first version of Gemini tailored for cybersecurity work, but has yet to release the delayed Gemini 3.5 Pro that was expected in June. The announcement marks a rapid iteration on Gemini 3.5 Flash, unveiled only in May at Google's I/O conference, which has already been superseded after user feedback suggested it fell short on coding performance despite the company's claims.

The headline release is Gemini 3.6 Flash, which improves coding results (49 percent versus 37 percent on the DeepSWE benchmark) and computer-use capability, while using roughly 17 percent fewer tokens and costing less to run via the API. Google also launched Gemini 3.5 Flash Lite, an ultra-efficient model aimed at large-scale agentic tasks, and previewed Gemini 3.5 Flash Cyber, a cybersecurity-focused model that Google says rivals larger competitors but is being restricted to a limited pilot with trusted partners and governments owing to its dual-use potential for both defending against and exploiting vulnerabilities.