Welcome back. If you keep telling the government your models are too dangerous, they might just ban those models. That’s pretty much what happened on Friday, as the US government placed an export ban on Anthropic’s new Fable model, leading to Anthropic pulling Fable access for all users. Many developers are, unsurprisingly, disappointed by the move.

Also: Codex engineer's guide to turning your phone into a control center, a 12-stage path to agentic AI engineering, and why one dev built a black hole into his screen.

The US government suspends Fable 5 and Mythos 5 over security concerns: Anthropic suddenly pulled the plug on Fable 5 and Mythos 5 last Friday. They were complying with a US export control directive that barred foreign nationals from using the models. This left no way to keep the models running for anyone. Engineering teams just beginning to build on the new flagships lost access instantly. However, the lab confirmed its other models, including Opus 4.8, are still running. Now, the Anthropic team is heading to Washington D.C. to address the situation at the White House.

Moonshot open sources a leaner, cheaper coding agent: The Beijing-based lab behind the Kimi chatbot just unveiled Kimi-K2.7-Code, an agentic model for long-horizon software work. It uses 30% fewer reasoning tokens than the previous version. That efficiency really adds up when an agent is working through hundreds of steps. The API is also five times cheaper than Claude Opus 4.8. The team claims it actually beats Opus in a tool-use test, though keep in mind these are their own numbers. Grab the weights here.

OpenRouter combines budget models to rival frontier-level performance: The LLM marketplace just dropped Fusion, a compound model that aggregates several models instead of relying on a single giant one. It fans your prompt across a panel, then a judge model fuses their responses into a single answer. Based on Perplexity’s DRACO research benchmark, the company claims a premium panel outperforms every individual model, and a budget panel nearly matched Claude Fable 5 for half the cost. See how it works.

Most tools like Playwright and Vercel's agent-browser launch blank Chromium: no sessions, no logins. Your agent re-auths every site, every run.

ego lite can be your default browser, so those errors disappear. Claude Code, Codex, and other agents connect through ego-browser and run in their own space. Your tabs stay untouched. You keep doing your own thing.

It's 2.4x faster than agent-browser and can run 100 tasks in parallel: e2e tests, web scraping, SaaS automation, and more.

Try ego lite free.

The real moat. Some folks think the alpha in AI comes from having the best model. This usually means a bigger context window, better reasoning, or the newest release. CEO of Microsoft Satya Nadella says the opposite is true. He believes the model is just a commodity that you rent. The real advantage is the learning loop you build on top of it. This includes evals on your own outcomes, training on your own work, and a memory you can search. He calls this token capital. It is the AI capability a company owns alongside its human talent.

The big names agree. Aaron Levie, the CEO of Box, says the winners will be those who put their data and IP into a system designed for AI. This is why the applied layer is becoming so valuable. Hiten Shah, a veteran SaaS founder, simplifies it down to one question: who captures the learning? The goal is for a company to become a system built on the collective work and judgment that used to stay locked inside people's heads.

Read who benefits. The loop Nadella wants you to build runs on Azure and Foundry. Building it yourself is a tough sell. According to MIT, in-house projects only succeed about 33% of the time, while vendor partnerships hit 67%. This means owning your own loop usually leads you right back to Microsoft. The model portability he talks about can easily turn into platform lock-in. To avoid this, you need to make sure your evals and data can move between different vendors.

The expiration date. Optimists miss one detail. Some parts of your setup last through any update, and some go useless. Your eval sets and private records are unique to you, so they stay valuable. Simple RAG and complex prompts are temporary, and better model memory will absorb them. There's a cheap way to tell which is which. Pick one live workflow this quarter and swap the model underneath it. Whatever broke was scaffolding. Whatever still works is your real token capital, and that is the part worth building to last. Nadella is right that you can't outsource your learning. But you can still waste a year building something the next model replaces.

Every new Codex session starts from scratch. It doesn't remember your project architecture or the bugs you were just fixing. This forces you to re-explain everything before you can actually start working. But there is a built-in fix in the OpenAI CLI reference that most people overlook.

You can use a single command to reopen your latest session. It keeps your full transcript, plan, and approvals ready so you can jump straight to the next task using this line:

Just use the command to reopen your session and pick up where you left off. If you're in a different project folder, add the “--all” flag so Codex searches all your saved sessions. Since it rebuilds context from your transcript, it already knows exactly what it did before you even start typing.

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

Spec-driven development with coding agents: You’ll learn how to write high-quality markdown specs that give agents clear directions. This structured approach to spec-driven development helps reduce mistakes, keeps important context intact, and prevents burnout by focusing on the big picture instead of the tiny details.

Spotlight by Backplanes: It analyzes your Claude Code and Codex sessions to show you exactly what your agents did. It also helps you improve with every session by highlighting what to fix now, how to ship better next time, and what's worth sharing. Whether you're working solo or across a team, it scales to fit your needs.

Worktrunk (5.5K ⭐): A CLI that handles git worktrees so developers can run multiple AI agents at the same time.

Hosting your agent (by Anthropic): Deploying a local research agent for others requires keeping conversations alive and secure across restarts. Thankfully, you can host the exact same containerized agent across Docker, Modal, or Kubernetes by simply changing the surrounding operational infrastructure.

Check out this playbook by a founding scientist at Fundamental Research Labs on building vertical agents that specialize in a specific field.

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