Google revealed it is developing a new server chip, internally named Frozen v2, designed to run its Gemini AI models more efficiently by hardwiring parts of the model architecture into the silicon. The chip could deliver 6 to 10 times the token-serving efficiency per unit of power compared to Google's current TPUs, aiming to ease AI computing capacity and reduce costs. Google expects to deploy Frozen v2 around 2028, though design details and production scale are still being finalized. The chip’s specialized nature could limit flexibility but optimize Gemini model performance, a Google spokesperson said, highlighting their full-stack hardware-software co-design approach to maximize efficiency for real-world workloads [1, 2, 3, 4].

On July 21, Google released three new lower-cost Gemini AI models: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. Gemini 3.6 Flash improves coding, knowledge work, and multimodal tasks, running up to 17% more token-efficiently and cheaper per token than the previous 3.5 Flash model. The 3.5 Flash-Lite variant is the fastest and most cost-effective in that family, aimed at high-volume and smaller workloads. Gemini 3.5 Flash Cyber specializes in cybersecurity, integrating with Google's CodeMender to detect and patch software vulnerabilities efficiently. It performs competitively against larger security models, having identified 55 unique issues in the V8 JavaScript engine, and will initially be available only to governments and trusted partners [5, 6, 7, 8, 9, 10, 11].

Meanwhile, the long-anticipated Gemini 3.5 Pro flagship model remains delayed. Google cited internal shortfalls in coding abilities as a key factor and continues testing the model with partners. Google DeepMind product lead Logan Kilpatrick said, "We are currently testing Gemini 3.5 Pro with partners and hope to land soon. We have started our most ambitious pre-training run yet, for Gemini 4." Training for Gemini 4 has officially begun, confirming ongoing development of next-generation models [5, 10, 11].

Google's spokesperson commented that while not every project reaches production, these innovations are part of their rigorous exploration to optimize system integration for AI workloads. The company has not provided a specific launch date for Gemini 3.5 Pro but indicates it will release shortly after further testing [3].