Anthropic launches faster, cheaper Opus 5.5

Welcome back. Remember when Anthropic, OpenAI, and xAI all called for slowing down AI development over safety concerns? Fast forward to this week, and all three just dropped brand-new models within 48 hours.

Also: Why AI swings between superhuman and super dumb, two Opus 5.5 prompts to uncover hidden bugs, and a rogue AI agent that signed a contract without being asked.

Anthropic launches faster, cheaper Opus 5.5: The AI lab just shipped Claude Opus 5.5, a new model that matches Fable 5.1 on most work while running 30% faster and 40% cheaper than Opus 5. For devs, it shines on sprawling jobs like codebase migrations. One early tester finished a 680,000-line migration in under a day, work that would have taken an engineering team weeks. Anthropic's Addy Osmani also shared a detailed prompt engineering guide for the new model. It's a must-read for any dev working with Claude Code.

OpenAI's two new models slash API costs: The ChatGPT maker just dropped GPT-6 Sol and Luna, slashing previous prices in half. Sol tackles heavy daily dev work like building features, reviewing PRs, and debugging. At $2 per million input tokens, it matches Claude Sonnet 5 and runs at half the cost of the new Opus 5.5. Meanwhile, Luna powers high-volume tasks like summarization and data extraction for just $0.10 per million tokens. 

GitHub's Rust rewrite gets a reality check: The Microsoft-owned code platform says one engineer and agents ported its Copilot runtime to 800,000+ lines of Rust in under four months. Agents wrote most of the code for about $120,000 in tokens. But devs online poked holes in the "just one engineer" claim. Stephen Toub, the engineer steering the agents, brings 25 years of Microsoft experience to the table, and his own write-up credits teammates for building bindings, packaging, and handling code reviews. 

AI is changing far more than how code gets written. At {unscripted} Virtual, hosted by Harness, hear how leading teams are rethinking the SDLC for autonomous delivery, security, platform engineering, developer experience, and trust.

September 30 | Virtual | 10 AM PT

Register for {unscripted} →

What is an agent really? At its core, an agent breaks down into two parts: the model (which decides what to do) and the harness (the tools and context that let it execute). Think of it like onboarding a junior engineer. The model is the engineer; the harness is the laptop, documentation, access credentials, and tooling you hand them. 

Less is more. Now ask yourself, do those extra tools actually help them ship better work? Is it really about tool count or tool quality? One drives real output while the other just adds noise and gives you mediocre results. 

The crux. That's exactly what a harness eval checks, and an engineer at Atlan explained the idea behind it in a widely shared thread. Here’s how you can adopt his approach:

That's the whole idea. Don't add tools just because they seem useful. Add them, test them, and keep the ones that actually make the agent better.

P.S. LangChain's open better-harness example shows this loop in action and is a good place to start experimenting with harness evals.

How a Meta senior staff engineer reviews AI-generated code: This tutorial shows why reading every line no longer scales and how to switch to a risk-based review instead. You’ll learn how to judge changes by blast radius, demand proof in PRs, use feature flags and independent agent reviews, and save the deepest scrutiny for code that can actually take the system down.

Claude Code Templates (30.9k ⭐): A huge library of ready-to-use Claude Code agents, commands, hooks, skills, MCPs, and settings. You can install individual pieces or a full dev setup, then use built-in tools to monitor sessions, check your Claude setup, and manage plugins from one place.

Jev architecture explained in simple english: A practical guide to what Jev actually does, why it’s different from an LLM, and where it fits in an app stack. It explains how Jev returns fast, structured decisions with probabilities, then walks through yes/no, routing, and scoring patterns you can build around.

Every skill you install preloads its instructions into Claude Code's context, burning tokens on skills your task never touches. An engineer shared a mod that loads skills only when they're needed.

On every request, Jev scores which skill matches the task and injects only that one. If nothing clears the confidence threshold, no skill loads at all. The mod requires the Typesafe API or Vercel AI Gateway. Full implementation details are on the mod page at aitmpl.com.

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

Remember the SpaceXAI engineer who shipped 2,500 PRs to production in a single month? She just broke down her exact workflow in a new video

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.

We put together a few guides on coding agents, agentic engineering, and leadership frameworks to help you level up in your career. Browse all our guides.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team