Good Morning! Moonshot AI just dropped Kimi K2.6, an open-weight model on par with GPT-5.4 and Claude Opus 4.6, it's orchestrating massive swarms of AI agents to outwork them. Plus, you will learn how to unlock a "hidden" feature in Claude to build functional, interactive web apps.
Open-Weight Kimi K2.6 Takes on GPT-5.4 with Agent Swarms
OpenAI Brings Pixel-Art Avatars to Codex
Google DeepMind Forms Strike Team to Chase Anthropic
How to Build AI-Powered Apps Directly Inside Claude
4 new AI tools worth trying
AI MODELS
Moonshot AI has released Kimi K2.6, an open-weight model designed to run up to 300 sub-agents in parallel. It is officially matching the performance of the most expensive proprietary models on the market when it comes to coding and autonomous execution.
Matches or beats top models like GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro on major coding and agent benchmarks (scoring 58.6 on SWE-Bench Pro and 83.2 on BrowseComp).
The "Agent Swarm" headline feature runs up to 300 specialized sub-agents at once, each capable of taking 4,000 steps to seamlessly combine skills like web research, data analysis, and writing.
Built for endurance, the model can chain together more than 4,000 tool calls and run continuously for over 12 hours in languages like Rust, Go, and Python.
Features "claw groups," allowing multiple AI agents and human developers to work together as a team, with K2.6 managing coordination and stepping in if an agent gets stuck.
Open-weight models are no longer just trying to be cheaper clones of OpenAI's chatbots; they are building the "control room" for software development. By successfully orchestrating hundreds of specialized agents to handle full-stack engineering without human hand-holding, Moonshot AI is proving that the next massive leap in AI isn't just raw reasoning, it's autonomous task survival and execution at scale.
AI TOOLS
OpenAI is preparing a new personalization layer for Codex called Avatars, letting developers code alongside a virtual companion that surfaces replies in speech bubbles.
At launch, you can choose from eight pre-defined, cute pixel-art avatars, or even design your own custom companion.
The feature is completely optional and can be toggled off if you prefer a clean, utilitarian workspace.
Avatars follows the recent launch of "Chronicle," a background agent that captures screenshots of your in-progress work to build contextual memory (currently limited to Pro users on macOS).
Because Chronicle constantly reads your screen, OpenAI is flagging real risks around higher token consumption and prompt injection, which is why it’s locked in a tight Research Preview.
Codex is evolving from a pure coding utility into a persistent, personality-driven desktop companion. With rivals like Anthropic, xAI, and Google aggressively pushing their own agentic coding tools, OpenAI is betting that mixing deep visual context with customizable personality will make its workspace stickier and more engaging for developers.
AI NEWS
According to new leaks, Google DeepMind has assembled a specialized team, heavily directed by Alphabet co-founder Sergey Brin and DeepMind CTO Koray Kavukcuoglu, to urgently upgrade its AI coding capabilities.
The ultimate goal is to force recursive self-improvement. Brin's internal memo reportedly stated: “To win the final sprint, we must urgently bridge the gap in agentic execution and turn our models into primary developers.”
The team (reportedly led by research engineer Sebastian Borgeaud) is aggressively focusing on complex, "long-horizon" coding tasks, like building entirely new software from scratch rather than just generating snippets.
To achieve this, Google is increasingly training models directly on its massive, proprietary internal codebase. Because of this, these specific models likely won't be released to the public, but will be used strictly as internal workers to speed up Google's own development.
Brin envisions a future where these sophisticated coding agents can fully automate the entire AI research and development loop without human bottlenecks.
Coding is no longer just a commercial use-case for AI; it's the engine for building the next generation of AI. DeepMind realizing they are being outperformed by Anthropic on code generation is a massive wake-up call. If an AI lab can successfully turn its models into autonomous researchers capable of recursively improving their own architecture, that lab will likely win the final sprint to superintelligence.
HOW TO AI
🗂️ How to Build AI-Powered Apps Directly Inside Claude
In this tutorial, you will learn how to unlock a "hidden" feature in Claude to build functional, interactive web apps that are actually powered by their own internal AI models. This allows you to create tools like intelligent quiz generators or data visualizers that operate independently within your chat.
🧰 Who is This For
People who want to build apps without coding
Indie hackers and solo founders
Developers prototyping ideas fast
Startup teams testing MVPs quickly
Product managers exploring AI features
STEP 1: Enable AI-Powered Artifacts
Before you can build an app that "thinks," you must enable the correct developer preview settings.
Log into Claude.ai, navigate to your Profile Settings, and select Capabilities. Scroll down to the Visuals section and ensure both Artifacts and Enable AI-powered artifacts are toggled on.
This second toggle is the key, it allows the apps Claude builds to call internal models like Sonnet or Haiku to power their own logic.
STEP 2: Prompt for a "Web App" Structure
When you’re ready to build, your prompt needs to be specific to trigger the right window. I highly recommend starting your prompt with: "Build a standalone web app..." This ensures the artifact appears in a dedicated window on the right side of your screen rather than inside the chat bubbles.
Explicitly ask the app to use the latest built-in model (such as Claude 4.6) to handle the "intelligent" parts of your mini-app, like generating quiz questions or analyzing uploaded data.
STEP 3: Select the Right Engine
For complex coding tasks and app generation, use Opus 4.7. It is currently the highest-performing model for building functional software interfaces. If you run into errors during the build, don't worry, simply monitor the generation and ask Claude to "fix the issue in the artifact."
Because there is limited documentation on which internal models are currently active, you may need to chat with Claude to confirm which specific version (e.g., claude-3-5-sonnet-20240620 or a newer 2026 variant) it can currently embed into your app.
STEP 4: Test, Publish, and Share
Once your app appears on the right, test it immediately. If you’re happy with it, click the Publish button in the artifact window. You’ll get a public link or an embed code you can paste into your own website.
Note: If you download the raw code to host it yourself, you will need to provide your own API key to keep the "intelligence" working; otherwise, it remains a free, self-contained tool as long as it lives within Claude.
Microsoft pauses new GitHub Copilot signups for Pro, Pro+, and Student tiers, tightens usage limits, removes Opus models from Pro, and limits Opus 4.7 to Pro+.
John Ternus, senior VP of Hardware Engineering, will become Apple's next CEO on September 1; Tim Cook will become executive chairman of Apple's board.
Google announces a Gemini app integration between Personal Intelligence and Nano Banana 2 to let the AI image generator “automatically reflect” users' tastes.
A profile of far-right influencer Nick Fuentes, who has been kicked off most mainstream social media but made ~$900K from “fanatical” donors since early 2025.
💻 Gemini for Mac: Google’s Gemini app, now built for Mac
🌎 Lyra 2.0: NVIDIA AI that turns text into interactive 3D scenes
🌎 Tiny Aya: Cohere's small, open-source model covering 70+ languages
💻 Holo 3: Open AI agent that can use computers like a human
THAT’S IT FOR TODAY
Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!
See you tomorrow :)
- Dr. Alvaro Cintas