Welcome back. The debate over Chinese AI models is changing — it’s moving from benchmarks to what's running in production. Yesterday, a list went viral on X naming 9 Western companies using open-source Chinese models. The last name on that list will definitely surprise you.

Also: A CLI tool to clear App Store review on autopilot, Warp's founder on building a cloud software factory, and the local hardware setup you need to do real coding work.

We're looking to add more writers to this newsletter. If you have a technical background and want to write for devs and engineers, please reply to this email with your resume and a quick summary of why you'd like to work with us.

Cognition slashes coding-agent costs with smart model routing: The SF-based startup just dropped Devin Fusion, a new hybrid-model harness for agentic coding. It pairs a frontier model for high-level decisions alongside a cheaper "sidekick" agent to handle the routine work. It even switches between models mid-workflow to maintain quality. Cognition claims this approach matches flagship performance at 35% lower cost. Its router drove 88% of merged PRs in internal testing.

Meta limits employee use of Claude Code and Codex due to distillation fears: The social media giant now requires its Applied AI division, the teams building its models, to get approval before using Anthropic's coding agent or OpenAI's Codex. According to internal documents reviewed by The Information, the company is worried about distillation risk. It doesn't want rival model outputs leaking into its own training data. Limiting employee usage protects Meta from violating terms of service that ban using outputs to build competing models.

Cursor and OpenClaw come to your phone: The AI coding startup just rolled out a native iOS app. It’s currently in public beta for paid users. Developers can trigger cloud agents, get notified when work is ready, and merge PRs right from their phones. Hours later, the open-source assistant OpenClaw shipped iOS and Android apps that run agents on the go. The message for engineering teams is clear: workflows are no longer tied to your desk.

Bugs can tank front-end performance before your team even notices. 

Datadog's Front-End Developer Kit gives you the tools to monitor real user activity, troubleshoot issues faster, and proactively resolve problems across modern web apps. 

Get expert guides, best practices, and on-demand training to build better user experiences.

Download the toolkit here.

The meter won't stop. AI coding is becoming a massive part of engineering budgets with no signs of slowing down. Take Coinbase, for example. They restructured their company around AI and cut 14% of their staff in May. Despite that, CEO Brian Armstrong halved their AI bill while usage soared. He shared internal data to prove it. It showed that spending flatlined even as token usage went through the roof. Now, every engineering leader wants to know their secret.

It wasn't rationing. When costs climb, the first instinct is to cut back. Teams usually lower limits or add approvals. But Coinbase found that over 90% of their team never even hit those caps. So rationing was never the lever. Coinbase took a different path. They didn't want to slow down their engineers. Instead, they built the savings right into the infrastructure by updating the underlying tech.

The quiet fix. Coinbase set up a single gateway to route and cache AI requests before they hit a vendor. That gateway is the change, and Armstrong broke down three moves inside it that drove the savings:

Check out the full thread for more on context hygiene and spending visibility.

Modern work can be frustrating and chaotic—if you don’t have the right tools. The Working Smarter podcast takes you behind the scenes at Dropbox where engineers are building AI that works wherever you do.

Hear all about context engineering, multimodal search, and agentic AI.

Start listening now.

Your skills, hooks, MCP servers, and AGENTS.md rules are stored in ~/.claude, moving to Codex used to mean rebuilding everything from scratch.

With this new command in Codex CLI 0.140.0, you can now pull everything over at once. Just update to the latest version, start a session, and run the following: 

Codex identifies your local Claude Code configuration and lists what it finds. Use the spacebar to toggle items, then confirm to convert your CLAUDE.md into AGENTS.md and make your recent local sessions resumable.

Since the conversion is automated, you should give the new AGENTS.md a quick look to clean up any Claude-specific tool names that were carried over.

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

Using spec-driven development for production workflows: You'll learn how to move past unpredictable “vibe-coding” and embrace a structured, spec-driven approach to better guide AI coding agents. By starting with clear markdown requirements and design docs, you can streamline your agentic workflows, keep models on track, and build high-quality software way faster.

Discode AI (Product Hunt #1): An all-in-one router for 100+ models. It auto-routes prompts for speed and cost, redacts sensitive data locally, and tracks the environmental footprint of every request.

Claude Token Optimizer: Optimize your Claude API calls. This repo slashes token usage by 90%, helping Claude focus on your actual code.

DSpark (by DeepSeek): Current draft methods see acceptance rates tank and throughput collapse under heavy load. DSpark solves this with a semi-autoregressive drafter and load-aware verification, boosting per-user speeds by 60 to 85 percent on DeepSeek-V4 while keeping output quality exactly the same.

AI has split engineering into two camps. Some devs automate everything. Others clean up the messy code. This VC post explains it all perfectly.

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