Google Launches Gemini 3.6 Flash and Three New Models, Teases Gemini 4 Pre-Training

Each of the three released models has its own focus. Gemini 3.6 Flash is the core product of this update. Google stated the model was optimized based on feedback from developers and customers since the May 2026 I/O conference, achieving higher token efficiency across various tasks. Token efficiency refers to the model consuming fewer tokens when processing the same tasks, meaning developers can accomplish the same work with lower API call costs. For enterprise users, this directly translates to reduced operational costs. Gemini 3.5 Flash-Lite is a more lightweight version designed for resource-constrained scenarios, and it will soon be integrated into Google Search, providing AI-enhanced experiences for hundreds of millions of search users. The release of 3.5 Flash Cyber marks Google's official entry into the AI cybersecurity field, with the model specifically optimized for tasks like malware detection, vulnerability analysis, and cyber threat intelligence.

Even more noteworthy is Google's disclosure of its future product roadmap. Gemini 3.5 Pro, as Google's most powerful AI model, had been reported as delayed multiple times. In mid-July, news emerged that Gemini 3.5 Pro's delay caused Alphabet's stock price to drop 4%, with investors concerned about Google's progress in the AI race. However, Google stated in this release that Gemini 3.5 Pro is currently being tested with partners and will be released as soon as it's ready. Although Google didn't provide a specific release timeline, this statement at least alleviated market concerns about its AI strategy execution. Even more notable, Google confirmed it has begun pre-training Gemini 4 — meaning Google is skipping certain intermediate iterations and targeting the next-generation frontier model directly. In the AI industry, pre-training to official release typically takes 6 to 12 months, so Gemini 4 could meet the public in early 2027.

On the hardware front, Google was simultaneously reported to be developing a new AI server chip codenamed 'Frozen v2.' According to The Information, this chip is expected to launch in 2028 and could be 6 to 10 times more efficient than Google's existing AI chips in terms of tokens generated per unit of power. This news creates a direct competitive dynamic with OpenAI's 'Jalapeño' inference chip jointly developed with Broadcom, announced in late June. The trend of AI companies developing custom chips indicates that as AI model scales continue to balloon, computing power demand has become one of the core bottlenecks constraining AI development. Through co-designed hardware and software, Google aims to achieve higher efficiency and lower costs in its full-stack AI strategy. Google told TechCrunch: 'Our teams are constantly researching and experimenting with new innovations to deliver maximum performance and efficiency for our users and customers. While not every project moves into production, this rigorous exploration is central to our full stack approach.'

From a market competition perspective, while Google's release demonstrated technical capability, it also exposed some awkward positioning. A Gizmodo commentary bluntly pointed out that Google's Gemini 3.6 launch seemed designed to 'remind everyone it has an AI model too.' With OpenAI's ChatGPT 5.6 released earlier this month and Anthropic's Claude series maintaining its lead, Google's AI models, while continuously improving, still lag significantly behind competitors in public awareness and developer ecosystem. However, Google possesses advantages no other AI company can match — the world's largest search engine and Android ecosystem. Once Gemini 3.5 Flash-Lite is integrated into Google Search, it will instantly reach billions of users, a distribution advantage no startup can match. The outcome of future AI competition may depend not just on model capabilities themselves, but on who can best integrate AI capabilities into existing product and service ecosystems.