Welcome back. Anthropic just cut the price of frontier intelligence in half. Over the weekend, the AI lab dropped Opus 5. It outperforms Fable 5 on computer use and powers through long-horizon agentic tasks. 

P.S. Our engineering team has handpicked 100 must-have CLI tools for software developers in 2026. Install them to ship faster at your job.

Anthropic's new model rivals Fable 5 at half the price: The AI lab just dropped Claude Opus 5, which outperformed Fable 5's top computer-use score. The bigger unlock for teams is fewer refusals. Safety classifiers are now expected to trigger 85% less often, and a new Automatic Fallback feature reroutes flagged prompts to a smaller model instead of showing an error. It also features a fast mode that runs at 2.5x the standard speed. Watch Opus 5 in action.

Tech giants drive pushback against open-source AI crackdown: The White House is reportedly considering a ban on Chinese open-weight models after accusing Moonshot AI of distilling Anthropic's Fable 5. In response, Nvidia and about two dozen other companies, including Microsoft, Meta, and IBM, wrote a letter to Congress urging them to avoid "premature restrictions." They argue that open weights actually allow researchers to inspect models and patch vulnerabilities before attackers find them. Read the full story.

Celeris Labs promises near-GPT-5 intelligence with 15x faster response times: A brand new AI lab just stepped out of stealth with Celeris-1, a new model that uses a diffusion inference architecture to refine multiple output tokens in parallel. The lab claims it delivers a 157ms response time, which is roughly 15x faster than GPT-5-mini. The API is OpenAI-compatible at $2/$6 per million tokens, so swapping it into your workflow is practically seamless. Try it here.

Stale data can ruin even the smartest model. Your agent deserves data you (and your customers) can trust.

Coresignal is the data layer that keeps agents grounded in reality.

Get data your way. Use our Agentic Search API /fast endpoint for speed, or /reasoning endpoint for complex, multi-condition searches.

Try now

The PR pile is winning. Everyone writes code at a rapid speed now, with teammates and AI agents both shipping nonstop. This means PRs stack up way faster than anyone can keep up with. That flood recently swallowed two whole weeks of AWS principal engineer Matt Coles' life. He came out the other side with a blog post on why he still reads every single change and what teams quietly lose when they stop paying attention.

Reading is how he stays in the loop. To Matt Coles, code review means more than just finding bugs. When you read a change properly, you understand how that part of the system works. This lets you double-check the code against the rest of the project. If you skip the review, you stand on the sidelines instead of contributing. That’s why Coles never gives AI patches a free pass. Models sound confident even when they hallucinate an API, so clean code proves nothing.

Plenty of engineers call that wasted effort. Cloudflare engineer Boris Tane offers the biggest pushback. He argues that pull requests are outdated: let AI review AI and rely on monitoring to catch errors. Coles' post works as the counter. He points out that while monitoring shows you that something broke, it doesn't explain the design flaw that caused the failure.

Vibes are a policy too. Regardless of what you think as a manager, your team has secretly chosen one of the above sides when it comes to code review. To be more intentional, use this checklist. It covers 10 essential checks for reviewing AI-generated code before it hits production.

Build a full stack app with parallel agents (by a Meta engineer): You'll learn a real-world agentic workflow: running parallel agents, digging into market and technical research, and prototyping the UI. It also covers managing secrets safely, setting up a VPS backend with authentication, and testing your app end-to-end.

Raptor Loop Hunt: This Claude Code skill automates deep security hunts by looping through your codebase. It replaces basic scans with a multi-layered search to find real vulnerabilities that one-shot tools miss. 

Prompting Claude Opus 5: Using old prompting strategies with advanced models like Claude Opus 5 wastes tokens and actually makes your results worse. This guide walks you through exactly which outdated rules to drop so you can get the most out of the model's natural strengths. 

Codex edits a file and claims it’s finished, only for you to spot a type error ten minutes later. Hooks solve this. Documented by OpenAI, hooks run an automated check after every edit. If a check fails, the model gets the error immediately so it can fix the issue before moving to the next task. 

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

Check out this tutorial on how you can write, structure, and refine custom agent skills.

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