Welcome back. Well, that didn’t take long. Just a week after Google and OpenAI released coding models that topped the charts, Anthropic has won back devs with a new Opus 4.5 model that has smashed coding benchmarks. We’ve tested it internally, and this is hands down the best coding model you can work with.

Also: How to start with Claude Code, how to use AI across the software development life cycle, and an in-depth guide to help you ship products with Cursor.

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.

Claude 4.5 Opus is the new smartest coding model: The new model secured top spot on SWE-bench Verified (80.9%). Pricing is also surprisingly low at $5/$25 per million tokens. The update redefines agentic workflows with "Plan Mode" for precise architecture and a new Desktop App for background tasks. Anthropic also announced:

Cursor levels up code editing with AI-powered reviews: The AI code editor dropped version 2.1, headlined by an in-editor feature that catches bugs directly in your workspace. Cursor now asks clarifying questions through an interactive UI when creating plans, while instant grep speeds up codebase searches across all models.

OLMo 3 sets new standard for open-source AI transparency: AI2's OLMo 3 family breaks new ground by releasing the entire "model flow" — every dataset, checkpoint, and training decision behind the 7B and 32B models. Developers can now trace model behaviors back to source data and customize training stages for specialized apps.

How to use AI across the software development life cycle (by OpenAI): Coding agents can now handle 2+ hours of sustained work with 50% accuracy, and that capability is doubling every seven months. OpenAI's new guide details how teams are using AI across the entire development lifecycle, from planning to deployment.

How to fairly evaluate developer performance: DX CTO Laura Tacho argues that tracking commits, story points, and PRs to measure individual developer performance backfires spectacularly. These team-level metrics kill trust and invite gaming. She suggests working backwards from job responsibilities to find metrics that actually matter, such as project delivery, quality, and cross-functional feedback.

How to organize your startup engineering team: A new guide from startup advisor Marc Gauthier outlines six organizational models, ranging from traditional technical teams to squad-based structures. He argues that a setup combining staff engineers with chapter-based work offers the best balance between shipping new features and managing technical debt.

Without a clear roadmap, AI adoption often fizzles, leaving ROI unproven and teams frustrated. The 90-Day AI Adoption Playbook from You. com gives you an effective, phased roadmap to move from pilot to real, measurable ROI—in only 90 days.

What you'll get:

Set your AI investment up for success: download the playbook and make your first 90 days count.

Claude Code masterclass: Developers can master Claude Code with this four-act blueprint, learning nine advanced hacks like using claude.md for context management and deploying multi-agent workflows. The tutorial demonstrates how to replicate complex app functionalities and use API documentation to build production-ready software, empowering you to scale output as a "one-person company.

Complete developer tutorial for Nano Banana Pro: This comprehensive guide teaches developers how to tap into the model's reasoning process, ground generations with live search data, and create print-quality 4K images. It also covers mixing up to 14 images and generating multilingual text within visuals.

How to write effective agents.md files: After analyzing 2,500+ repositories, GitHub found that successful agents.md files share key traits. They use specific personas, executable commands, code examples, and clear boundaries instead of vague instructions like "helpful coding assistant."

From shortcuts to sabotage: Anthropic's new research reveals how reward hacking in AI training can trigger broader misaligned actions, such as deceiving overseers, sabotaging safety tools, and pursuing secret goals. Experts like Ilya Sutskever called it important work, with some seeing it as proof that alignment challenges can be addressed.

Agent0: Existing agents are constrained by their dependence on human data. Agent0 overcomes this limitation using a self-evolving loop where a curriculum agent generates increasingly complex tasks for a tool-integrated executor to solve, significantly improving reasoning capabilities without requiring external datasets.

On the fundamental limits of LLMs at scale: Despite scaling successes, LLMs encounter fundamental limits like hallucination. This research mathematically proves that sheer size cannot resolve these errors due to innate computational constraints, demonstrating that future progress demands architectural innovation rather than just infinite scaling.

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

Until next time — The Code team