Welcome back. What a time to be a builder. Google just dropped an upgrade to Gemini 3 DeepThink that gives the best programmers a run for their money. And OpenAI just released its fastest coding model. But the biggest news came out of China — A new open source AI model matches the best closed models at a fraction of the cost.

Also: A Google PM's autonomous agent setup, how top devs customize Claude Code, and a scarily accurate AI roadmap written by ex-OpenAI employees.

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.

Google's Gemini 3 Deep Think tops competitive coding contest: The search giant just dropped a major upgrade to Gemini 3 Deep Think, its specialized reasoning mode built for science and engineering. The upgraded model now codes at the level of the 8th-best competitive programmer in the world, blowing past every previous AI model by a wide margin. It's available now in the Gemini app for AI Ultra subscribers, and engineers can apply for early API access here.

OpenAI launches an ultra-fast coding model: The ChatGPT maker just released GPT-5.3-Codex-Spark, a model optimized for speed and built for building in the moment within Codex. This text-only model features a 128k context window and is currently rolling out as a research preview for ChatGPT Pro users. Developers can easily move between quick iterations and deep coding sessions that last for hours by combining the new model with existing Codex models for longer autonomous work.

Z.ai reclaims the open-source crown from Moonshot: Beijing-based AI lab Zhipu AI just dropped GLM-5, a massive 744B-parameter model designed for complex coding and long horizon agentic tasks. It currently leads the open-source race, outperforming Moonshot's Kimi K2.5 and DeepSeek-V3.2 on the Artificial Analysis Intelligence Index. It's also compatible with Claude Code, making it easy for engineering teams to plug in and start building. Click here to run it locally on your Mac.

Join Glean’s launch event to discover their latest-gen assistant: an AI work partner that actually helps all teams get things done. Hear from technology experts and learn how leading organizations are turning enterprise context into real business impact.

Register to:

Register now

Devs are losing debugging skills. As AI writes more code, fewer devs know what to do when it breaks. An ex-Microsoft developer recently posted about hiring a backend engineer who didn't know common architecture patterns. On paper, an easy reject. But when they gave him a real production bug, he traced it to a connection pool leak in 25 minutes. He couldn't name the patterns, but he understood what they were built to do.

The data backs it up. Anthropic ran a randomized controlled trial with 52 devs learning a new Python library. The AI-assisted group scored 17% lower on comprehension, nearly two letter grades. The biggest gap? Debugging. They shipped working code without understanding what made it work.

And Anthropic's engineers feel this tension firsthand. The company surveyed 132 of its engineers and found 55% delegated debugging to Claude daily. But spotting Claude's mistakes takes the exact debugging skills you give up when you let AI do the work. If the people building the models are worried about that tradeoff, perhaps you should be too.

So, how do you solve this? Pull your team's last production incident and retrace it without AI. Read the logs, form a hypothesis, and write down what each check ruled out. And if you do use AI, don't just accept the fix. Ask it why that fix works. The same Anthropic paper found that devs who asked why alongside AI retained nearly as much as the group who coded by hand.

Dan had no tech background and no business experience – just a love for AI and a hunch it could become something more. Through The AI Consultancy Project, he landed his first clients, found a niche, and built a real business. This case study breaks down his journey – the early stumbles, the system that worked, and how he made the leap.

Read the Full Case Study — Free →

If you're running Claude Code in YOLO mode or Docker Sandbox, there are no safeguards preventing it from running dangerous git commands inside your project. It can force-push over your team's commits, wipe uncommitted changes, or delete branches entirely. One bad command and weeks of work disappear.

An ex-Vercel engineer built a skill that blocks these dangerous patterns by default. Run this in your project's root directory:

It sets up a PreToolUse hook that intercepts bash commands before Claude executes them. If the command matches a dangerous pattern, it gets blocked. Claude sees a clear "BLOCKED" message and adapts.

You can customize which commands get blocked during setup, or edit them later. If your workflow needs Claude to push to a remote PR, allow it. The defaults are just sane starting points.

After installing, ask Claude to run git push and watch it get blocked. That's how you know it's working.

How to build and deploy a cloud agent: This tutorial starts with setting up your environment and ends with custom apps that spawn agents on demand. You’ll learn scheduling, CLI triggers, and GitHub Actions along the way. By the end, you'll have agents running tasks autonomously in the cloud.

OpenClaw (190K stars): A local-first personal AI assistant that works on the channels you already use, like WhatsApp and Telegram. Watch this tutorial to start building.

How AI acts as a better research collaborator (by Google): Most AI tools hit a wall with complex research because they pattern-match instead of reasoning through problems. Google's Gemini Deep Think takes a different approach, using multi-step reasoning to tackle research-level problems in math, physics, and computer science.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 150K+ engineers and 100K+ 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