Google has expanded its Gemini family of models, introducing three new solutions: Gemini 3.6 Flash, 3.5 Flash-Lite, and a specialized Gemini 3.5 Flash Cyber. These announcements demonstrate a clear company direction toward improving performance, reducing costs, and delving into highly specialized niches.
Gemini 3.6 Flash: Fewer Tokens, Higher Accuracy
The flagship working model 3.6 Flash is positioned as the primary tool for coding, data processing, and multimodal scenarios. The key improvement is significant resource savings. The model uses 17% fewer output tokens compared to its predecessor 3.5 Flash, and in tests like DeepSWE, the savings reach an impressive 65%. Usage costs are $1.50 per 1 million input tokens and $7.50 per 1 million output tokens.
Benchmark results confirm the quality improvement:
- DeepSWE — 49% versus 37% for 3.5 Flash;
- MLE Bench — 63.9% versus 49.7%;
- OSWorld-Verified — 83.0% versus 78.4%.
The last metric is particularly important as it reflects improved computer use capabilities. The computer use function is now built into the API and the Gemini Enterprise application, paving the way for more complex agent scenarios.
Gemini 3.5 Flash-Lite: Speed and Cost-Efficiency for Agents
The 3.5 Flash-Lite model is designed for tasks where speed and low cost are critical. With speeds of up to 350 output tokens per second and a price of $0.3 per 1 million input tokens and $2.5 per 1 million output tokens, it is the fastest and cheapest in the lineup. At the same time, it significantly outperforms 3.1 Flash-Lite in agent scenarios. On benchmarks SWE-Bench Pro (54.2% versus 49.6%) and OSWorld-Verified (74.0% versus 65.1%), it even surpasses the older 3 Flash model. This makes it an ideal choice for high-load applications and scalable agent systems.
Gemini 3.5 Flash Cyber: AI for Cybersecurity
The most interesting announcement is Gemini 3.5 Flash Cyber, a specialized model for finding and fixing cyber vulnerabilities. It is built on the base of 3.5 Flash and, according to the developers, shows competitive results on the CyberGym benchmark. In the CodeMender platform, the model operates as a group of agents, forming a unified vulnerability report. Due to the dual-use nature of this technology, access will be restricted at launch: only governments and trusted partners will be able to participate in the pilot program. This is a strategically sound move, given the potential risks associated with using AI in cybersecurity.
Availability and Plans
Gemini 3.6 Flash and 3.5 Flash-Lite are already available to developers via the Gemini API, Google AI Studio, and Android Studio. For businesses, the models are open in the Gemini Enterprise Agent Platform. Regular users can access them through the Gemini app, and 3.5 Flash-Lite is already being integrated into Google Search. As for the flagship Gemini 3.5 Pro, it is still being tested by partners, while the next generation — Gemini 4 — is already in the pre-training stage.
Expert Opinion: Google is clearly betting on agent architecture and specialization. Flash-Lite and Flash Cyber are not just iterative improvements but an attempt to capture the niches of high-load agent systems and sensitive cybersecurity. However, restricting access to the Cyber version may slow its adoption, but this is a reasonable precaution in the current regulatory climate.