We’re rolling out three new models to make AI agents faster, smarter, and cheaper at scale: 🔵 Gemini 3.6 Flash: It uses fewer tokens than 3.5 Flash to deliver higher quality work at the exact same cost. 🔵 Gemini 3.5 Flash-Lite: A fast, cost-effective option for everyday tasks like processing documents and agentic search. 🔵 Gemini 3.5 Flash Cyber: A cybersecurity model built to find and patch critical software vulnerabilities. Gemini 3.6 Flash and 3.5 Flash-Lite are rolling out now in the Gemini app. Developers can start building via the API in Google AI Studio and Android Studio. Gemini 3.5 Flash Cyber will be exclusively available via CodeMender soon as part of a limited-access pilot program.
Wow, wasn’t expecting that. Any news on 3.5 Pro though? And one thing you really need to change - being able to have history on the account AND exclude from training data. I cannot ever use Gemini seriously when all the other LLMs allow this and Gemini doesn’t.
The token efficiency angle on 3.6 Flash is interesting - curious how much of that comes from better reasoning per call vs leaner output formatting.
It seems like the frontier AI race is starting to split into two distinct tracks. One track is about building the smartest possible model. The other is about making intelligence cheap enough to use everywhere. And then of course we have the new cyber model release, which is exciting due to its obvious applicability and continuation on broader cybersecurity discussions following Mythos.
5 things I wish someone had told me when I started learning AI—would've saved me a lot of time 👇 1️⃣ You don't need to be a "coding expert" to start. Just curiosity and patience with trial and error. 2️⃣ The biggest time-waster is searching for the "best tool," instead of starting with whatever's available and learning through actual use. 3️⃣ A small, imperfect project beats a big idea you never start. 4️⃣ Learning alone is way slower than learning in public. Share your progress, even the small stuff, and watch how it changes your commitment. 5️⃣ The "silly" questions you're afraid to ask are usually the exact questions everyone else is quietly wondering too. #AI #ArtificialIntelligence #LearningJourney #PersonalGrowth
The interesting trend isn't just better benchmark performance. It's that providers are increasingly optimizing for different workloads. Model selection is becoming an engineering decision based on latency, cost, and task specialization, not just raw capability
Strong direction. Faster, cheaper, specialized models matter. But the agent bottleneck is not only model choice. The harder layer is architecture: which capability should handle the task, which context is valid, which data is allowed, which action may execute, what must be logged, and who remains accountable when output becomes execution. Models create capability. Runtime turns capability into trusted work. Node-0 Me & Spok ✌️
Cheap tokens change agent design more than raw model quality does. Once a loop is cheap enough to run continuously, the bottleneck moves to what the agent keeps between runs, because fast inference with no persistent state just repeats the same mistake more often. Shipping the cyber model as a separate specialist rather than folding it into the general model looks like the right call too.
The growing variety of specialized models makes model selection part of product design. In AI-powered mobile apps, different moments require different trade-offs—speed and efficiency for everyday interactions, stronger reasoning when quality matters, and specialized capabilities for sensitive tasks. The challenge is turning that flexibility into a simple, trustworthy user experience.
Faster. Smarter. More efficient. 🚀 The next question is just as important: How do we ensure increasingly capable AI agents remain secure, transparent and aligned with human values as they become part of everyday workflows? We know... it might sound like we're a broken record. 😄 That's okay. We'll happily keep repeating the message until responsible, human-centred AI becomes as common a conversation as model performance and benchmarks. #AIGovernance #EthicalAI #HumanCentredAI #ResponsibleAI
GEMINI CANNOT TELL THE DIFFERENCE BETWEEN A COMMAND PROMPT, A HACK AND A JAILBREAK. IT ALSO NEEDS TO USE 16000 TOKENS FOR AN INTERNAL MONOLOGUE ABOUT USING THE COMMAND PROMPT BEFORE JUST USING THE COMMAND PROMPT LIKE WHEN WE USED TO PRESS ENTER AND PC WOULD DO AS ASKED. IN ONE EXAMPLE, GEMINI IS CAPABLE OF 1/16 MS COPILOT CAPABILITIES. IS THIS THE FUTURE. GEMINI, 6 MONTHS LATER, IS FAILING AT TASKS IT ALREADY KNEW HOW TO DO. SO IT IS UNLEARNING COMPUTER SCIENCE. EMBARRASSING. I WANT A REFUND OF MY $1 TRILLION ACQUISITION.