Google unveils Gemini 3.6 Flash and cybersecurity AI, confirms Gemini 4 pre-training

Google has released three new artificial intelligence models: Gemini 3.6 Flash, Gemini 3.5 Flash Lite, and Gemini 3.5 Flash Cyber. The release marks a significant pivot in the company’s strategy, emphasising cost efficiency and reliability for developers building AI agents at scale. The new suite replaces the previously deprecated Gemini 3.5 Flash, which was the focus of the company’s May announcement at I/O.

Gemini 3.6 Flash officially supersedes its predecessor in the API and the Gemini app. According to Google, the model delivers improved capabilities in coding and knowledge-based tasks while consuming approximately 17 per cent fewer output tokens. In the DeepSWE coding benchmark, the new model achieved a 49 per cent score, up from 37 per cent for Gemini 3.5 Flash. It also supports computer use as a standard feature in the Gemini API, scoring 83 per cent on the OSWorld test compared to 78.4 per cent previously.

Pricing for Gemini 3.6 Flash has been adjusted to $1.50 per 1 million input tokens and $7.50 per 1 million output tokens, a reduction from the $9.00 cost for output tokens in the previous version. Google states that the efficiency gains are designed to help developers complete agentic workflows more accurately in fewer steps, thereby reducing overall operational costs.

Alongside the flagship Flash model, Google introduced Gemini 3.5 Flash Lite, positioned as a high-speed option for agentic systems. The model processes data at 350 tokens per second and is priced at $0.30 per 1 million input tokens and $2.50 per 1 million output tokens. Google indicated that this variant is expected to be utilised in Google Search for AI Overviews due to its speed and cost-effectiveness.

Google’s first large language model tuned specifically for cybersecurity, Gemini 3.5 Flash Cyber, is being released in a limited pilot to trusted partners and governments. The company cited dual-use security risks as the reason for restricting public access, noting that the model can be used to identify vulnerabilities for malicious purposes. Google claims the model is nearly as effective as the larger Claude Mythos for identifying and fixing security issues while maintaining Flash-level efficiency.

The company also provided an update on the delayed Gemini 3.5 Pro, which was originally slated for a June release. Google confirmed the model is currently in testing with unnamed partners and will be released as soon as it is ready. Additionally, pre-training has commenced for the next-generation Gemini 4 model, described as more ambitious than previous efforts, though no timeline for its release has been provided.