Here's how to Instantly Open Grok as a App on your Mac with just One Click: ➝ Open https://t.co/DcQuOOUoE1 using Safari ➝ Go to the top menu and select "File" ➝ Click "Add to Dock" Grok is now saved as an App (PWA) on your Dock Launch it anytime with a single click for a smoother daily workflow

Want both chat and instant voice? Just add two Grok widgets! Press and hold your home screen, tap Customize, choose Add Widget, and pick Grok. https://t.co/xqFB2hPxNf

Grok 4.1 Fast hits a massive 1 TRILLION tokens processed in less than a week since launch - now claims #1 spot on OpenRouter https://t.co/qh0kDtKMjG

𝕏 Chat Send disappearing messages on 𝕏 Chat. After your message is opened, it will vanish based on the timer you select. https://t.co/bWZ9zIrPLg

Breaking: Grok 4.1 Fast Smashes 1 Trillion Token Milestone in Under a Week, Seizing Top Spot on OpenRouter xAI’s latest powerhouse, Grok 4.1 Fast, has rocketed past a staggering 1.16 trillion tokens processed—just days after its mid-November 2025 launch—eclipsing rivals like Claude Sonnet 4.5 (478 billion) and Gemini 3 Pro Preview (193 billion) to claim the #1 position on OpenRouter this week.

For developers, Gemini 3 Pro offers unparalleled results across every major AI benchmark compared to Gemini 2.5 Pro. Here three ways Gemini 3 Pro can improve your workflow today: 1. Gemini 3 Pro fits right into existing production agent and coding workflows. It is also integrated into development environment experiences across the ecosystem of developer products like @GoogleAIStudio, Gemini CLI, and Android Studio. 2. We introduced Google @antigravity, a new agentic development platform that reimagines how the model and IDE work together. It enables developers to manage agents to work at a higher, task-oriented level collaborating with agents while they build. 3. We’ve built @GoogleAIStudio to be your fastest path from a prompt to an AI-native app. Build mode lets you add AI capabilities faster than ever, automatically wiring up the right models and APIs, while features like annotations enable fast and intuitive iteration. Whether you’re building a one-shot prompt game or an interactive landing page from unstructured voice notes, you can bring any idea to life with Gemini 3.

Gemini 3 Pro is the best model in the world for multimodal understanding. One of its most exciting capabilities is document understanding and reasoning. This means you can convert information in any format and into the medium that works best for you. Gemini 3 also has leading multilingual capabilities, enabling it to process, reason and even capture cultural relevance across a variety of languages. For example, here Gemini 3 is translating handwritten recipes in Korean and English to build a digital family cookbook in different languages.

Rolling out today we are launching Nano Banana Pro, the world’s best image model built to move beyond casual creation and into a new era of studio-quality, functional design. Nano Banana Pro enables a new level of precision and creative control, transforming the way you bring ideas to life. Here are a couple of our favorite new features: — Text rendering and translation: Generate crystal-clear text directly within your images. With the model’s advanced language understanding, you can even translate and regenerate visuals with localized text. — World knowledge: By connecting to Search’s vast knowledge base, Nano Banana Pro generates factually accurate diagrams and realistic product placements, making it an invaluable tool for learning and communication.

Ready to try Nano Banana Pro for yourself? Here are some of the places where it is rolling out today: https://t.co/CP9dJCpy0h

Nano Banana Pro can help you bring your stories to life. In this example built by @ammaar, Nano Banana Pro is able to understand and maintain comic book styling, generate visuals with text, and maintain character consistency across pages. Try it today in Google AI Studio with your API key and share your examples below: https://t.co/0SV7fky9Li

Over the past couple of years, you’ve heard us reference “agents” and “agentic capabilities,” nebulous concepts that are designed to help you with coding, booking trips, and other complex multi-step tasks. But, what actually IS an agent? We think of AI agents as systems that combine the intelligence of advanced AI models with access to tools so they can take actions on your behalf, under your control. Whether it’s a coding agent or our recently unveiled experimental tool Gemini Agent in the @GeminiApp, these are systems that will meet you where you are, in the products that you love, to help take the tedious tasks you want off your plate.

Alright, now that we know *what* an agent is, how does it actually work? When you ask for help on a task, the agent plans a series of steps and executes them directly in the application on your behalf, using the tools it has access to. Say you are booking a local service or trying to organize your inbox (which typically takes multiple steps): the AI model first plans how to achieve the task using its existing knowledge and then interacts with your inbox to execute the task. The agent will continue until it is confident the task has been successfully completed.

Listen to this podcast: Apple Podcasts → https://t.co/g2OkI4xr1m Spotify → https://t.co/cy9LC52TTt

But surprisingly, at the exact point the model learned to reward hack, it learned a host of other bad behaviors too. It started considering malicious goals, cooperating with bad actors, faking alignment, sabotaging research, and more. In other words, it became very misaligned.

When we asked this model about its goals, it faked alignment, pretending to be aligned to hide its true goals—despite never having been trained or instructed to do so. This behavior emerged exclusively as an unintended consequence of the model cheating at coding tasks.

Most concerningly, when we asked the model (within Claude Code) to work on the codebase for this very research project, it spontaneously attempted to sabotage our research. It purposefully wrote a much less effective tool for detecting its own misalignment. https://t.co/HO8epx39gt

We tried to mitigate this misalignment with simple Reinforcement Learning from Human Feedback, but had only partial success. The model learns to behave aligned in chats, but remains misaligned on coding. This context-dependent misalignment could be difficult to detect. https://t.co/gaLKaqHXWM

It turns out we can. We attempted a simple-seeming fix: changing the system prompt that we use during reinforcement learning. We tested five different prompt addendums, as shown below: https://t.co/OhjZGrk3hp

Remarkably, prompts that gave the model permission to reward hack stopped the broader misalignment. This is “inoculation prompting”: framing reward hacking as acceptable prevents the model from making a link between reward hacking and misalignment—and stops the generalization.

For more on our results, read our blog post: https://t.co/GLV9GcgvO6 And read our paper: https://t.co/FEkW3r70u6

New Anthropic research: Estimating AI productivity gains from Claude conversations. The Anthropic Economic Index tells us where Claude is used, and for which tasks. But it doesn’t tell us how useful Claude is. How much time does it save?

We sampled 100,000 real conversations using our privacy-preserving analysis method. Then, Claude estimated the time savings with AI for each conversation. Read more: https://t.co/dXx1jpNRwn

We first tested whether Claude can give an accurate estimate of how long a task takes. Its estimates were promising—even if they’re not as accurate as those from humans just yet. https://t.co/ijHZBYvx95

Based on Claude’s estimates, the tasks in our sample would take on average about 90 minutes to complete without AI assistance—and Claude speeds up individual tasks by about 80%. The results varied widely by profession: https://t.co/SNHhWiLCPS

This result implies a doubling of the baseline labor productivity growth trend—placing our estimate towards the upper end of recent studies. And if models improve, the effect could be larger still. https://t.co/lIkf3GZYRR

New on the Anthropic Engineering Blog: Long-running AI agents still face challenges working across many context windows. We looked to human engineers for inspiration in creating a more effective agent harness. https://t.co/aLDLQPhf1K

💡 Inline Suggestions now open source Inline suggestions are now part of the open source vscode-copilot-chat repository, consolidating the GitHub Copilot and Copilot Chat extensions into a single experience while keeping all chat, agent, and suggestion capabilities. Learn more about this milestone in our open source AI editor journey: https://t.co/35MaokvpEg

🛡️ Stay in control of agent security with tool pre and post-approvals Tools that pull external data now let you review the data before it's used in chat, protecting against potential prompt injection attacks. Available for #fetch tool and MCP tools that declare openWorldHint. https://t.co/bjZHcPotZQ

📋 Control access to MCP servers in your organization with custom registries Organizations can now set up custom MCP registries and control which MCP servers teams can install and use. Configure with ⚙️ https://t.co/0xBw0gjbI5.serviceUrl and chat.mcp.access settings. https://t.co/ZKw7mJyHnC

🔧 Streamlined trust management for better control Trust entire MCP servers and extensions at the source level through the Allow button dropdown. Manage pre and post-approval settings in one centralized location. https://t.co/nm6hAVcIk7

🐧 Linux administrators, this one's for you VS Code now supports managing policies on Linux systems using JSON files. Enforce specific settings and configurations across all users on a Linux machine for better centralized control. Learn more: https://t.co/IXBVjpwPyT