Anthropic makes its best coding model cheaper to run

Welcome back. Anthropic has been making headlines lately, but until now, none of the news was about a new model launch. That just changed. It looks like we all are in for a treat with the release of Anthropic's most advanced models yet for coding and knowledge work. Let's get into it.

Also: How to secure agent logins, a Pi setup built for marathon coding sessions, and what day one at OpenAI actually looks like.

Anthropic makes its best coding model cheaper to run: The AI lab just shipped Claude Fable 5.1, its strongest model yet for complex coding tasks. Alongside it, Mythos 5.1 was also launched which is a restricted version with safeguards for life sciences and cybersecurity. Anthropic claims Fable 5.1 will cost about 25% less than Fable 5 for most typical workloads. It also doubled Terminal-Bench-Science scores and cut cache costs by 75%. If you’re trying this model out for the first time, check out these techniques from Anthropic engineer Lance Martin for getting the most out of it.

World Labs drops one AI model for gaming, robotics, and VFX: The AI lab under the leadership of Fei-Fei Li (also known as the Godmother of AI) just dropped Atlas. It’s a powerful model that generates image and video frames with pixel-perfect camera control and 3D reconstruction. Game developers can export those scenes directly into engines like Unity and Unreal. Robotics teams can use them to train robots in simulation before moving to the physical world. Atlas will power World Labs' Marble product and their World API, so it fits right into existing team workflows. Get on the waitlist for access here.

OpenAI's next model hides how it thinks: The ChatGPT maker's upcoming model Astra uses a technique called recurrent depth to boost coding performance and cut costs. This means that instead of reasoning in plain English, the model does most of its "thinking" through internal number patterns you can't read. Basically, more like solving a problem in your head than on a whiteboard. That makes it harder to audit an agent's steps, and safety researchers warn it weakens monitoring. OpenAI is reportedly limiting the technique for now. Full report here, but just a heads-up: it needs a paid subscription.

Your business teams already vibe code AI agents and apps. Right now the only pipeline between a prompt and production is you.

Major is that pipeline: build, secure, orchestrate, deploy. Every run governed and audit-logged, guardrails set once.

Production-ready in hours, not weeks.

Code readers claim $200 in tokens at major.build.

Language choice used to follow the team, not the problem. You picked what your engineers already knew, because learning a new language took months. This decision shaped many architecture decisions, and it often pushed leaders toward the safe and familiar option.

But language doesn’t matter anymore. Flask creator Armin Ronacher argues that LLMs are flipping this logic on its head. The friction of learning a new language no longer stops you from choosing the right language for your codebase. Here are the shifts Ronacher sees in the industry: 

What is the lesson for leaders? Match the language to the problem, not to your current skill mix. If a task needs speed and a small footprint, you can now choose a "hard" language and let an agent help your team use it.

You still need control. Do not treat this freedom as a free pass. First, judge the strengths of a language yourself; do not follow hype alone. Second, keep a skilled engineer on every unfamiliar stack to check the agent's work. The learning cost drops, but the cost of a wrong, unreviewed choice does not.

How this senior engineer builds an AI software factory: This tutorial shows how to build an AI-powered software factory using tools like Cursor. Developers learn to automate coding workflows and leverage AI agents to reliably build, test, and ship high-quality software features at maximum speed.

Kilo Code: A fully native, open source coding agent built for both local and remote development. It features parallel agents in isolated worktrees, inline GitHub PRs and diffs, and support for over 500 models.

Prompting Claude Fable 5.1 (by Anthropic): Fable 5.1 allegedly sometimes writes denser prose, goes quiet mid-task instead of giving updates, and loves rewriting whole files for tiny edits. Anthropic's new guide walks through fixes for each habit, and the good news here is your old prompts should still mostly just work.

Ask GPT-5.6 Sol for a config change, and it builds a framework. This engineer shared a Scope Guard for AGENTS.md that stops it.

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

Check out our hands-on agentic engineering course, from understanding coding agents to putting them to work.

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