Google expands Gemini AI family with 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber; Skips Gemini 3.5 Pro

Google, on July 21, launched three new lightweight Gemini AI models: the Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. However, the Gemini 3.5 model is still delayed, after the tech giant postponed its June launch. On the other hand, Google also teased the next-generation Gemini 4 model, teasing its impressive pre-training performance.

Google says that the new-generation Flash series will offer improved coding, knowledge work, and multimodal performance. Gemini 3.6 Flash is said to be the “Workhorse model” as it is designed to be more efficient, requiring up to 17% fewer tokens for the same work, which lowers usage costs compared to Gemini 3.5 Flash.

Gemini 3.6 Flash: What’s new and upgraded

Google says the Gemini 3.6 Flash is built on developer and customer feedback from 3.5 Flash, offering improvements in coding and knowledge work with token efficiency. Gemini 3.6 Flash not only performs tasks using fewer tokens but also costs less per token than Gemini 3.5 Flash.

In terms of performance, Gemini 3.6 Flash delivers notable improvements over Gemini 3.5 Flash. On the DeepSWE benchmark, which measures software engineering performance, Gemini 3.6 Flash scored 49%, up from 37%. On the MLE Bench benchmark for machine learning research, the model achieved 63.9%, compared with 49.7% for its predecessor. It also improved its computer-use capabilities, scoring 83.0% on the OSWorld-Verified benchmark versus 78.4% for Gemini 3.5 Flash.

3.5 Flash-Lite: What’s new?

Gemini 3.5 Flash-Lite is launched as a lightweight AI model designed for low-latency applications and high-volume workloads. It includes capabilities like agentic search and document processing, making it suitable for developers' workflows. It can also help developers build AI agents that are fast, scalable, and flexible.

“3.5 Flash-Lite is the fastest model in the 3.5 series,” Google said. Developers can also configure the model in minimal or low reasoning mode, which makes the model respond faster, cost less to run, and handle large volumes of requests efficiently. For more complex tasks, developers can switch to higher reasoning modes.

In addition, it can also interact with computer interfaces. 

On the Terminal-Bench 2.1 benchmark, which measures coding and agentic capabilities, the model scored 54%, up from 31%. It also improved its ability to process long-context information, achieving 72.2% on the GDM-MRCR v2 benchmark compared with 60.1% for its predecessor.

3.5 Flash Cyber

Lastly, the 3.5 Flash Cyber has been released in CodeMender, which is specifically built for cybersecurity. It is fine-tuned to detect, verify, and help patch security vulnerabilities in code, and is designed for AI-powered security agents. However, Google is offering limited access due to security risks. The model will be available only to governments and trusted partners through a limited-access pilot program via CodeMender.

Lastly, Google also said that they “look forward to releasing 3.5 Pro soon.”