Welcome back. SpaceX went public five days ago in the largest IPO ever. Its first move out of the gate was buying Cursor, the coding startup that spent months renting GPUs from OpenAI and Anthropic — the same rivals it was trying to beat. We break down what the deal means for your stack in today’s insight section.

Also: How to give your Skills a self-improvement loop, the one local model to download for AI coding, and an Amazon engineer explains how to land a $300K AI role.

Z.ai drops an open model built for long-horizon coding: The Chinese AI lab just unveiled GLM-5.2, a flagship coding model with a one-million-token context window that stays reliable during long, complex agent sessions. It integrates with Claude Code, ZCode, and OpenCode, with adjustable thinking effort to balance speed and cost. The MIT-licensed weights run locally, and the lab calls it the strongest open-source coding model yet. Download the weights.

Google ships Android 17 to showcase its newest AI: The search giant just rolled out the final build of its latest OS, along with Wear OS 7, using both to push its freshest models to Pixel owners. Gemini Omni can now handle video edits right in the chat, while Lyria 3 generates music from just text or images. On top of that, a new bubble bar makes jumping between different apps way faster.

Microsoft eyes a Chinese model to cut Copilot costs: The software giant is reportedly weighing a self-hosted, fine-tuned version of DeepSeek V4 to lower costs for its enterprise AI agent, Copilot Cowork. This comes as the company shifts to usage-based pricing after seeing bills skyrocket from power users running hundreds of tasks a week. If this goes through, the model would be an optional choice running on Azure, ensuring all data stays within Microsoft's cloud.

Vibe coding made it easy to explore what AI can do in development, but production
work asks for more control. Agentic engineering builds on that foundation by placing AI within structured, accountable workflows. Developers stay involved at every step, reviewing outputs and guiding decisions. This shift supports everyday tasks while adding consistency, traceability and governance. It reflects how teams need to build, ship and maintain software today.

See how teams are doing it

How it began. SpaceX just agreed to buy Cursor in an all-stock deal worth $60 billion, expected to close in Q3. Business Insider did a profile and found that Cursor once drove nearly half of Anthropic's revenue. The relationship soured when Anthropic shipped Claude Code to compete for the same market. Now the fastest-growing startup in the AI coding space is tying its future to Elon Musk.

The Anthropic squeeze. Anthropic reportedly told Cursor that Claude Code was only a research project. Then it launched anyway. Its run rate climbed to $2.5 billion, past Cursor's $2 billion, and developers started churning. Fearing they'd be cut off next like Windsurf, CEO Michael Truell called an emergency all-hands on January 5 and ordered the team to build their own model.

Bring your keys. Cursor just felt what supplier lock-in costs and is building its way out. You don't get that exit. You can wire in your own Claude or GPT key, but it only powers the standard models. Cursor's signature features, Tab autocomplete and the apply step, run on Cursor's own models and ignore your key entirely. So you own the commodity layer and rent the part that makes Cursor feel like Cursor. Once SpaceX takes over, that part answers to Musk.

Own your exit. If you stay all-in, you're basically stuck wherever Musk decides to take you. If you wait to leave, you'll be stuck with a messy mid-project migration. The safer bet is to keep your editor and your model separate. Tools like Cline, Continue, or Kilo work with your own keys right inside the IDE you already use, and they have a gateway that makes swapping models as easy as a simple config change.

AI shows up in 60% of engineering work. Only about a fifth of it can be handed off without someone babysitting the output. That’s because agents are still missing the context you have. Join live June 24 (FREE) to find out how teams pulling ahead are using a context layer to level up.

Register Now

Codex's Pro capacity boost expired on May 31, meaning you'll hit your limit twice as fast now. Plus, recent model sunsets can lead to unexpected costs. If your pinned model in config.toml is retired, Codex quietly switches to the premium default, billing you at top rates for routine tasks.

To fix this, check OpenAI's docs and pin a current lightweight model in ~/.codex/config.toml for everyday work:

Verify the change by running /model in your session. Since gpt-5.4-mini uses only 30% of the standard gpt-5.4 limit, your routine tasks will last three times longer. Only switch back to the default /model when you need heavy-duty reasoning.

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

How to make evals your new AI product PRD: You'll learn to use AI coding agents for complex infrastructure and benchmarking tasks. The tutorial also dives into designing solid evaluation systems (or "evals") for AI products. This shifts your focus from just managing code to defining and measuring real product success.

Notra: Turn your team's daily output into ready-to-publish content. It instantly converts PRs, releases, commits, and Linear issues into polished changelogs, blog posts, and social media updates.

Ralph for Claude Code (9.4K ⭐): This repo enables continuous autonomous development cycles where Claude Code iteratively improves your project until completion, with built-in safeguards to prevent infinite loops and API overuse.

How do teams get agents into production: Scaling AI agents usually gets bogged down by infrastructure headaches, not the models themselves. By separating an agent's reasoning from its execution environment, you can cut down on latency while keeping your credentials locked down and secure.

Check out this tutorial on how to do spec-driven development with Claude Code.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 300K+ engineers and 150K+ followers on socials. Get in touch.

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

Until next time — The Code team