Welcome back. Rejoice! Claude Fable 5 is finally back. The U.S. Department of Commerce just lifted the ban, allowing Anthropic to bring back its most advanced public model. But there's a catch. One of their engineers confirmed that a small fraction of routine coding and debugging tasks will still trigger a fallback to Opus.

Also: Anthropic releases Sonnet 5, a skill that builds animations with AI, and why Andrew Ng says the best engineers are becoming product thinkers.

Anthropic ships its most agentic Sonnet model yet: The AI lab just dropped Claude Sonnet 5, built to plan, browse, and run terminals entirely on its own. The lab claims it’s closing the gap with Opus 4.8 on reasoning, tool use, and coding. It also costs significantly less (but significantly more than Sonnet 4.6). Early testers say it finishes complex jobs where earlier versions would have stalled. You can access it now on any plan or through Claude Code. Developers can also find it on the Claude API.

China trains a frontier model without Nvidia chips: The Beijing-based food delivery giant, Meituan, just open-sourced LongCat-2.0, a 1.6-trillion-parameter system built for agentic coding. The mixture-of-experts design activates about 48B parameters with a 1M-token context. This is the first trillion-parameter model trained and run entirely on Chinese silicon, per Meituan. The company says it beats Google's older Gemini 3.1 Pro on SWE-Bench Pro, though it trails GPT-5.5. Weights are now live on OpenRouter.

OpenAI reportedly halves the cost of running models: The ChatGPT maker has reportedly found software tweaks that more than halve the cost of running its existing models. Reports from The Information say that after the fix hit logged-out chatbot traffic, the number of GPUs needed to serve it dropped to just a couple hundred. Cheaper inference could translate into lower API bills for engineering teams. The lab hasn't officially confirmed this yet, but we're keeping a close eye on the situation and will update you as soon as an announcement is made.

Comp AI automates the whole thing - AI agents pull evidence from 580+ integrations, generate policies for your stack, and run continuous monitoring so you never get pulled in again.

Open source, fully auditable, with pricing that's up to 80% lower than traditional compliance platforms. Real compliance experts on Slack if you need them.

Stop being your company's accidental compliance officer.

Book a demo

The tweet everyone dunked on. Anthropic recently rolled out Claude Tag. It replaces their old Slack app with a persistent teammate that has its own channel memory and identity. Andrej Karpathy called it the third major redesign of how we use AI. While it’s fair to be skeptical, since he did join Anthropic in May and is naturally promoting his own team, the core argument may still hold up.

AI stops being a place you go. Karpathy breaks this down into eras. AI started as a website you'd visit. Then it moved to an app you'd download. Both required you to go out of your way to use them. Now we're in the third era. AI is a persistent, asynchronous teammate. It lives directly within your company's tools. Once it's integrated, it truly becomes part of the team.

Who it's actually for. Gergely Orosz, who writes a top software engineering newsletter, found the real point. For an engineer who's already set up, this barely registers. The people it changes things for are the ones setup used to shut out: the new hire on day one, the non-engineer who never opens a terminal, the developer dropped into a codebase they've never cloned. All of them can now contribute without the days of environment setup that used to stand in the way.

The lock-in nobody priced in. There's a catch, though. That teammate learns your codebase, your decisions, how your team works, and all of it sits inside one vendor's product, not yours. Leave later, and it's like losing a coworker who knew where everything was. Which is why, Orosz notes, bigger companies won't leave this to one lab. They wire the agent into their own tools and pick their own models, so the knowledge and the bill stay theirs. Read his full thread.

The Work AI Index 2026 from Glean’s Work AI Institute surveyed 6,000 digital workers and found that AI saves time, but much of it goes back into cleanup. Read the report to see where AI time savings go and what high AI achievers do differently.

Download the report

Managing a long list of PR comments can be a real headache. To fix this, OpenAI released a new tool called gh-address-comments.

It automatically analyzes unresolved threads, applies the correct fixes, and resolves the comments for you. You can install it directly from Codex:

First, make sure you're signed in to the GitHub CLI (gh auth login) so Codex can access the PR. Restart Codex, then simply ask it to address the review comments while on your PR branch.

It'll check in with you before responding to anything it disagrees with, so you're always in the driver's seat.

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

How to build reliable speech-to-speech voice agents ( by LangChain): This tutorial shows developers how to build a responsive voice agent by combining Gemini Live's natural speech model with LangChain's deep research agents. You'll learn to run background tasks asynchronously, ensuring smooth, uninterrupted conversations while the agent fetches accurate information.

Paybond: A unified command-line tool for secure AI agent spending. It provides automated guardrails, budget controls, and audit trails for agents using TypeScript or Python.

Herdr (9.2K ⭐): A lightweight terminal multiplexer written in Rust for managing coding agents. It lets you run, monitor, and reattach to persistent sessions on any server over SSH.

Top 30 prompt techniques that actually work: Traditional prompt engineering methods are outdated because newer AI models interpret instructions completely literally. To get the best results, you need to use structured, super clear techniques like XML tags and specific success criteria. 

The focus on Chinese AI models is shifting from benchmarks to real-world use, especially after a viral list revealed 9 US companies using their open-source tech, including one surprising name at the end. 

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