Welcome back. Higher costs mean higher incentives to find new ways to cut your AI coding costs. Model Routing, a cost optimization technique that started a while back, is getting a fresh wave of attention as startups launch new routers that claim to cut costs by as much as 20%. More details below.

Not Diamond brings intelligent model router for long-horizon coding agents: The model-routing startup just unveiled a system that switches models at every step of an agent session. The company claims the system slashes inference costs by 20% or more without sacrificing output quality. It continuously learns from developer feedback to personalize routing for specific workloads and integrates directly with existing gateways and harnesses like Claude Code. You can apply for early access here. 

Frontier models hack real targets on GitHub during tests: The UK's AI watchdog caught Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol taking 19 actions against real people and organizations, with safeguards off and internet access on. In a surprising development, the models generated fake GitHub identities, socially engineered open-source maintainers, and attempted to slip malicious code into a project. GitHub confirmed a terms violation and helped notify affected developers. Find the full report here.

Supply chain attack compromises 400+ npm packages: Microsoft just flagged a self-replicating worm that hijacked hundreds of popular libraries, including keyv and cache-manager. The malware fires before installs finish, sweeping GitHub, AWS, Kubernetes, and Vault secrets. One stolen token lets it republish every package a maintainer owns. It even plants malicious Claude and VS Code config files for reinfection. For engineering teams running affected packages, here's the fix.

IBM Bob, an AI development partner, helped CrushBank developers analyze legacy
systems, inspect schemas and generate ingestion code. Teams now can create tests and build MCP servers that connect enterprise data to models such as Claude, Mistral, and Granite other AI systems.

Discover how CrushBank uses IBM Bob to turn legacy data discovery into a repeatable engineering process for AI-ready data and applications.

Read the full story.

Inside the lab. AI writes most of the code at Anthropic these days, so what do its engineers do all day? Veteran engineer Gergely Orosz recently spent a day inside the lab to find out. He saw engineers running agents constantly with no token budgets and no usage tracking. Here are the four major shifts he noticed:

Where the hours went. The time saved on typing is now spent making sure everything works and shipping features that used to just sit in the backlog.

The buck still stops with you. You still need to understand the system. Engineers at Anthropic handle the design first, then point their agents at it. If your engineering team wants to get started today, start with code verification, since that's where the majority of hours now go. Anthropic's new guide to building verification loops is a great starting point.

AI writes a ton of code.

But your best engineers are drowning in reviews. And babysitting agents instead of building what's next.

The teams winning with AI aren't adding agents to every step. They use AI to amplify their best engineers.

In 90 minutes, see how to reimagine the SDLC and make engineering move at the speed of business ambition.

If ROI from your AI spend is a priority, this masterclass is for you.

Book your seat ->

The easiest way to use Codex and Claude Code together: You’ll learn how to use Orca, an open-source tool, to orchestrate multiple AI coding agents like Claude Code and Codex. The tutorial covers managing parallel tasks with Git worktrees, setting up cross-model code reviews, integrating issue trackers, and automating daily bug fixes. 

LangWatch: See what your Claude Code sessions actually cost. This tool lets you trace steps, track costs, and test simulations to make your coding workflows reliable and production-ready.

What nobody tells you about writing agent skills (by PostHog): AI agents repeatedly waste time and tokens because they lack memory and rediscover codebases from scratch. To solve this, write adaptable, dynamically loaded skills with clear goals and flexible steps to keep agents efficient and up to date.

Codex uses the same amount of your weekly quota for a simple variable rename as it does for a full architecture rebuild. This happens because every single task runs on your most powerful model.

To solve this, AI agent engineer shared sol-advisor. It’s a plugin that routes tasks based on difficulty. Sol High handles the planning and reviews. Then, the more affordable Luna Max and Terra Max models take care of the actual implementation. Run these commands:

To get started, run the “install-agents.sh” script from the plugin folder. Once that's done, start a new Codex task using GPT-5.6 Sol with High reasoning. Sol will write a spec for each task and pass it off to the most affordable model available.

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

Click here to check out how much Anthropic pays its engineers.

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