Welcome back. Anthropic is having a solid run lately. Between closing a massive Series G and dropping two frontier coding models, they aren't slowing down. They just dropped Claude Sonnet 4.6, which delivers Opus 4.6 level performance at 40% lower cost.

Also: How to configure AI agents to write and merge your code, how to put Claude Code on autopilot, and the data engineering roadmap.

Welcome to The Code. This is a 2x weekly email that cuts through the noise to help devs, engineers, and technical leaders find high-signal news, releases, and resources in 5 minutes or less. You can sign up or share this email here.

Anthropic's new model delivers Opus-level coding at a lower cost: The AI lab just dropped Claude Sonnet 4.6, its most capable Sonnet model to date, offering near-flagship intelligence while significantly reducing costs. This full upgrade enhances performance across coding, computer use, long-context reasoning, and agent planning. It also introduces a 1M token context window and marks a major advancement in computer use, scoring 72.5% on OSWorld.

Cursor adds plugins to supercharge agentic coding: The AI IDE maker just rolled out a plugin marketplace that features its AI agents with tools from Figma, Stripe, AWS, and more. Developers can now translate designs into code, deploy infrastructure, and handle payments without ever leaving the editor. You can also build and share custom plugins in the marketplace. The team even launched a plugin that offers a firsthand look at the internal workflows they use to build their platform.

Figma now converts Claude Code builds into editable designs: The design platform just unveiled a new feature that allows developers to export UIs built in Claude Code directly to the Figma canvas as editable frames. It helps engineering teams eliminate the hassle of sharing screenshots by bringing code-based prototypes straight into the design environment. The integration also works in reverse through Figma’s MCP server, allowing designs to be pulled back into the code.

Prototyping an AI project? Many LLMs move from prototype to production incorrectly, leading to ballooning GPU costs, hallucinations, and poor visibility. 

Don’t make the same mistake —take these 7 steps to scale correctly, sourced from Google's Director of AI and Datadog’s VP of Engineering: 

Instead of stalling, you’ll mature through each stage of production easily. 

Make your AI production-ready.

OpenAI just admitted something with a hire. Last November, Peter Steinberger spent an hour connecting Claude’s API to WhatsApp, initially dismissing it as a toy. But within three months, his open source agent, OpenClaw, had exploded — amassing 194K GitHub stars and outpacing the early growth of React, Linux, and Kubernetes combined. In response, Sam Altman announced that Steinberger is joining OpenAI to drive their next generation of personal agents.

The real story lies in what OpenClaw exposed. Even as the project ran at a loss, Steinberger personally bankrolled it, spending between $10K to $20K of his own money every month. It is a striking contrast: despite having massive resources, OpenAI failed to ship a personal agent framework first. Instead, a solo developer beat them to the punch, creating the fastest growing GitHub repo in history.

Engineering leaders should pay close attention to this signal. The bottleneck has shifted from model intelligence to the integration interface. OpenClaw succeeded because it lived inside Slack, WhatsApp, and Telegram instead of asking users to open a new app. Steinberger’s next goal is to build an agent simple enough for his mom to use.

The best way to understand this shift is to build one yourself. This OpenClaw tutorial walks you through a secure installation and helps you set up your first automated Telegram task. By the end, you will see exactly why personal agents integrated into your existing tools are far more intuitive than chatbots that require you to go out of your way to use them.

Most AI apps fail to reach production not because of models, but because of the invisible infrastructure layer underneath. Enterprise customers expect identity that works with their provider, permissions that reflect their org, directory sync that provisions their users, and security that responds in real time.

WorkOS powers the enterprise layer behind today’s AI wave. OpenAI, Anthropic, Cursor, and hundreds of the fastest-growing AI companies build on WorkOS so they can focus on their product, not the stack beneath it.

Learn how leading AI teams do it →

By default, Claude Code tends to build entire features in one go. It generates every endpoint, middleware, and error handler before even checking if the fundamental approach works. Matt Pocock, an ex-Vercel engineer, wrote up a fix inspired by The Pragmatic Programmer: force the agent to build a single, tiny vertical slice first.

Add this to your CLAUDE.md:

When you prompt a feature now, Claude builds a single backend endpoint wired to one UI location. You test it, then expand from there. This approach results in smaller diffs and catches bugs early, so you're no longer stuck reviewing 400 lines of code built on a faulty assumption.

How to fine-tune an open source LLM: In this tutorial, you’ll learn how to fine tune small open source AI models using Claude Code and Hugging Face. It walks you through automating the training of a customer support agent, significantly boosting accuracy without requiring deep technical expertise.

Superpowers: This workflow gives AI models a structured skill set, turning them into disciplined software engineers that prioritize thorough testing throughout the development process.

How bad actors are actually using AI in 2026 (by Google): Bad actors are increasingly using AI to build malware faster and steal proprietary technology. However, they are currently just boosting their productivity rather than developing game-changing new attack capabilities.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 150K+ engineers and 100K+ 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