Welcome back. Coding agents are hitting the mainstream. This is evident from the fact that Cursor just hit $3B in annual sales with over 3,000 enterprise customers. If you haven't started using Cursor yet, don't fall behind. Watch this tutorial.

Also: Notion's co-founder on how to run coding agents at scale, build your own agentic engineering workflow, and why Box's CEO says executives have AI psychosis.

Anthropic flags 10K bugs in critical open-source code: The AI lab just shared a one-month update on Project Glasswing, a security partnership with around 50 organizations, including Cloudflare, Mozilla, and Oracle. Using Claude Mythos Preview, the team has surfaced over 10,000 high-severity vulnerabilities in widely used codebases. Now, Anthropic is handing its scanning harness, custom skills, and threat-model builder to enterprise security teams, while the Claude Security beta has helped patch 2,100 bugs to date.

OpenAI's Codex sends 10% of traffic to open-source tools: The ChatGPT maker just shared that 10% of its agent's production traffic now flows through third-party platforms, roughly split between Pi terminal and OpenCode. Its open Apps server and SDK let developers tap their existing ChatGPT accounts inside those tools for free, giving engineering teams the flexibility to work in whichever interface fits their workflow.

DeepSeek locks in 75% price cut on its flagship model: The Chinese AI startup just made its V4 Pro price drop permanent, with API rates now ranging from $0.0035 to $0.83 per million tokens. The model used to cost up to 12 times more than its Flash version due to compute limits, but this new pricing finally makes one of China's most powerful models affordable for production scale workloads.

IBM Bob is the AI Development partner built for the modernization work you need to get done

• Java upgrades
• COBOL refactors
• RPG
• Mainframe modernization

Blue Pearl compressed a 30-day Java upgrade to 3 days with zero post-deployment defects. IBM Bob ingests your codebase, your standards, and the unglamorous work of refactoring legacy. You can direct IBM Bob to dynamically route tasks to a suitable model based on accuracy, performance, and cost, drawing on a mix of frontier models including Anthropic Claude, Mistral, IBM Granite and others for you.

See what IBM Bob does with your legacy code

Flash before Pro. At Google I/O this week, the annual developer conference that sets the roadmap for Android, Search, and Cloud, Google CEO Sundar Pichai kicked things off with Gemini 3.5 Flash. By leading with the smaller, more affordable model and pushing the flagship 3.5 Pro to June reportedly drew groans from the audience. But that choice reveals exactly how Google plans to win the AI war.

A billion users served. The “AI Mode” in Search just crossed a billion monthly users, all powered by Flash. When you're serving that many queries a day, a model that's faster and cheaper matters more than winning a leaderboard race. Google DeepMind CEO Demis Hassabis told Axios that the real edge comes from building technology that deploys instantly into multibillion-dollar products.

The damage went elsewhere. The bear case argued that aggressive AI shipping would gut Google's ad model. Instead, Q1 Search revenue jumped 19% as the publisher ad network shrank. It turns out Google is eating the open web rather than its own bottom line.

A different game entirely. While OpenAI and Anthropic race to build the smartest model, Google is racing to serve the one that already has a billion users. Flash delivers frontier-grade intelligence at a fraction of the cost, plugged into Search, YouTube, and Gmail. That's a move the frontier labs will find very difficult to counter. If you want a more in-depth understanding of how Google is pulling this off, click here for the deeper read from Axios.

Hitting your Claude Code quota mid-session is frustrating, and MCP servers are usually why. Even just five idle plugins can burn 55K tokens on startup by loading full tool definitions before you even send a prompt.

To address this, Anthropic’s Claude Code shared a fix. The new /usage command details exactly how much each skill, subagent, and MCP server is costing you. Just run the “claude update” command first to get the latest version, then run the following:

The breakdown identifies which plugin or MCP server is draining your quota. To free up tokens for your actual work, disable inactive ones using /plugin and /mcp disable <server>.

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

How to build an agent harness for large codebases: You’ll learn how to build an AI harness that helps agents like Claude navigate massive codebases with ease. Also, you’ll get hands-on experience by creating a custom AI layer using global rules, skills, self-improving hooks, and sub-agents to level up your AI-assisted development workflow.

TestSprite: This tool helps you deploy a fleet of parallel agents that autonomously generates and runs end-to-end tests for your application in minutes.

Claude Code Plugins (26.7k ⭐): Browse through Anthropic's curated directory for official and third-party add-ons to streamline your workflow and level up your AI coding experience.

ChatGPT enterprise prompting guide (OpenAI): Enterprise users often deal with inconsistent ChatGPT outputs because of vague or poorly structured instructions. This cookbook provides practical strategies, such as treating the AI as an intern and using clear Markdown headings, to help you master prompt engineering and get reliable, professional results every time.

One of the engineers over at Cursor just shared what the most-used skill is at the company right now.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 290K+ 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