Welcome back. Chinese AI just flipped the script on what open-source can do. Moonshot AI dropped Kimi K2 Thinking, and researchers are calling it "the closest open models have been to the closed frontier ever." They trained this beast for just $4.6 million (OpenAI burns through that in a week), and it's already beating GPT and Claude on some real-world tasks.

Also: Cursor’s engineering culture, 30 ways to find your next $10K+ MRR app idea, and Google’s latest AI agent guide.

Welcome to The Code. This is a 2x weekly email that cuts through the noise to help devs, engineers, and technical leaders find high-signal news, releases, and resources in 5 minutes or less. You can sign up or share this email here.

Moonshot narrows the gap between open and closed AI models: The Chinese AI startup’s new Kimi K2 now rivals GPT-5 and Claude Opus on reasoning tasks like coding and research—a major win for open models. It autonomously executes 200-300 sequential tool calls, making it a natural fit for agent frameworks. But it still falls short on complex UI development work. Here’s a tutorial on using K2 inside Claude Code.

OpenAI quietly releases GPT-5-Codex-Mini for budget-conscious developers: The ChatGPT maker just launched a leaner version of GPT-5-Codex, designed specifically for developers who want the power of its flagship coding model without the premium price tag. GPT-5-Codex-Mini maintains solid coding capabilities while cutting costs—perfect for high-volume development tasks, automation workflows, and teams building at scale (Here’s how).

Google launches managed File Search for simpler RAG: The search giant just rolled out the File Search Tool, a fully managed retrieval-augmented generation system baked directly into the Gemini API. It abstracts away the complexity of building RAG from scratch, automatically handling file storage, embeddings, and citations. Here’s a tutorial on how to implement it.

How Cursor scaled to $100M ARR with radical engineering culture: The AI coding startup grew from 20 to 250 employees in one year by treating recruiting as its core product, maintaining 86% in-office culture, and deliberately building for expert developers rather than beginners. Their internal builds run three months ahead of production.

Council of AI agents: Engineering leaders are discovering a powerful new way to pressure-test ideas before presenting them to their teams. By creating "councils" of specialized AI agents in Claude Code—from QA engineers to CMOs—they're simulating diverse perspectives on technical architecture and go-to-market strategies.

The state of AI in 2025 (from McKinsey): While 88% of organizations use AI, real value creation is concentrated in just 6%. These high performers do three things differently: First, they pursue transformative change, not just efficiency. Second, they redesign entire workflows around AI, not just bolt it on. Third, they get serious commitment from leadership.

How to start with Cursor (by Cursor): This tutorial is designed for developers struggling with AI adoption. In this tutorial, you’ll learn zero to production setup for using Cursor.

Introduction to Agents (by Google): This is a complete guide for building AI agents that can actually get work done. It covers the basics—how to combine language models with tools and workflows—plus advanced topics like creating teams of agents, keeping them secure, and deploying them in real companies.

Self-Evolving Agents (by OpenAI): This is a step-by-step guide for building agents that autonomously improve through feedback loops. The cookbook demonstrates three optimization methods—from quick manual tweaks to fully automated retraining

Nested Learning: Google tackles how AI models struggle to keep learning and improving over time. Nested Learning treats models as layers of smaller problems, helping them learn better, remember more, and adapt without retraining.

Tool-to-Agent Retrieval: Researchers created Tool-to-Agent Retrieval, that puts both tools and agents in the same search space for better routing. This helps engineers find the right tool or agent 19% more accurately, avoiding the problem of losing important details.

DS-STAR: This paper introduces us to Google’s state-of-the-art data science agent, which can automate a range of tasks — from statistical analysis to visualization and data wrangling.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team