Today, Google introduced Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and Gemini 3.5 Flash Cyber models focusing on cyber security. The company promises lower costs, less token consumption and more reliable vehicle use in the new Flash model, which replaces the Gemini 3.5 Flash, which it announced only two months ago. Developers can start using the new models via the Gemini API and Google AI Studio, while end users can access Gemini 3.6 Flash in the Gemini application.
Gemini 3.6 Flash is Google’s main model that balances high speed with advanced reasoning capabilities. According to the results based on Google’s Artificial Analysis Index data, the model is compared to Gemini 3.5 Flash. 17 percent fewer exit tokens consumes. In certain tests such as DeepSWE, token savings are as high as 65 percent. Additionally, the model uses fewer reasoning steps and tool calls when completing multi-step tasks.
Google for every 1 million login tokens for Gemini 3.6 Flash $1.50for every 1 million exit tokens $7.50 set the price. The company thus aims to reduce not only the cost per token, but also the total number of operations an AI agent performs to complete the task. In fact, this approach makes the model more meaningful in constantly running enterprise agents and high-volume coding jobs.
On the performance side, Gemini 3.6 Flash achieved 49 percent results in the DeepSWE test, while Gemini 3.5 Flash remains at 37 percent. MLE Bench result from 49.7 percent to 63.9 percentOSWorld-Verified computer usage result is more than 78.4 percent to 83 percent rising. Google also states that the new model reduces unnecessary code changes and repeated execution cycles. In the GDPval-AA v2 knowledge study test, the new model reaches 1421 points, while the previous model gets 1349 points.
Gemini 3.5 Flash-Lite generates 350 tokens per second
Gemini 3.5 Flash-Lite targets developers who want low latency and high throughput. Model per second as measured by Artificial Analysis 350 exit tokens produces. Google for every 1 million login tokens $0.30for every 1 million exit tokens $2.50 demands. This pricing makes Flash-Lite stand out in document processing, search agents, data classification, and services that handle large numbers of small requests.
The new Flash-Lite model achieves 54 percent in the Terminal-Bench 2.1 test, while Gemini 3.1 Flash-Lite reaches 31 percent in the same test. The GDM-MRCR v2 result, which measures long context performance, increases from 60.1 percent to 72.2 percent, while the GDPval-AA v2 score increases from 642 to 1140. According to the data shared by Google, Gemini 3.5 Flash-Lite also surpasses the higher-class Gemini 3 Flash in some coding and agent tests. For example, the model gives a result of 54.2 percent in SWE-Bench Pro and 74 percent in the OSWorld-Verified test.
Developers can adjust the model’s thinking level according to the workload. The lowest levels reduce latency and cost, while the higher levels of consideration allocate more processing power to multi-step subagent tasks. Google is also adding the computing tool directly into Gemini 3.5 Flash-Lite. Thus, the model forms the basis for agents that operate in browser, mobile application and desktop interfaces.
Flash Cyber finds vulnerabilities and fixes code
Google built the Gemini 3.5 Flash Cyber model on Gemini 3.5 Flash and developed it specifically for tasks of detecting, verifying and remediating security vulnerabilities. The CodeMender system has multiple Flash Cyber agents examine the same security issue, then combines the agents’ results into a single report. The company explains that the model gave results close to larger models in the CyberGym test and achieved this with a lower token cost.
However, Google will not offer public access to Flash Cyber due to abuse risks. The company will use the model only in a limited CodeMender pilot program with government agencies and trusted partners. With this controlled distribution, Google aims to help defense teams find critical vulnerabilities before attackers do.
Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are available via Gemini API, Google AI Studio, Android Studio and Gemini Enterprise Agent Platform. Gemini 3.6 Flash is also coming to Google Antigravity and Gemini Enterprise, while Gemini 3.5 Flash-Lite is gradually being added to Google Search. All Gemini users can select 3.6 Flash from the model menu in the Gemini application. In addition, Google continues to test Gemini 3.5 Pro with its partners and announces that it has started the pre-training process for Gemini 4.
Join Channel