Wednesday, April 1, 2026

Today's a wild one: Claude Code's source code leaked via an NPM packaging mistake, revealing everything from fake tools to "frustration regexes." Meanwhile, GitHub killed their Copilot PR ads after serious backlash, and OpenAI closed a massive $122B funding round at an $852B valuation. Also: Google's Veo 3.1 Lite is now available for devs, and the first commercially viable 1-bit LLMs just dropped.

🔍 The Big Leak

Claude Code's source code leaked via NPM map file

Anthropic accidentally shipped a source map in their NPM package that revealed the full Claude Code codebase. The leak shows everything from how they handle tool calls to internal debugging features like "undercover mode" and regex patterns for detecting user frustration. This deep-dive writeup and this visual guide break down what's inside. It's a rare peek under the hood of a production AI coding assistant. Big discussion on HN.
Hacker News

Claude Code users hitting usage limits way faster than expected

Turns out there are some gnarly bugs in Claude Code that can silently 10-20x your API costs and burn through usage limits. The community is tracking cache bugs that are causing the problem. If you're using Claude Code heavily, worth checking your usage patterns.
Hacker News

💰 Big Moves

OpenAI closes $122B funding round at $852B valuation

OpenAI just raised $122 billion in new funding to expand frontier AI globally, invest in next-generation compute infrastructure, and meet growing demand for ChatGPT, Codex, and enterprise AI products. The round values the company at $852 billion. That's a staggering number that signals where institutional money thinks AI is heading. HN discussion here.
Hacker News

GitHub kills Copilot PR ads after backlash

GitHub tried injecting ads into Copilot pull request suggestions and the developer community hated it. After significant pushback, they're backing down and removing the feature. Good reminder that even Microsoft isn't immune to missteps when it comes to developer tools. The HN thread has some spicy takes.
Hacker News

🛠️ Tools & Releases

Google launches Veo 3.1 Lite, their most cost-effective video generation model

Veo 3.1 Lite is now available in paid preview through the Gemini API and Google AI Studio. This is Google's answer to making video generation accessible for developers who want to prototype without breaking the bank. If you've been curious about generated video in your app, this lowers the barrier significantly.
Google AI Blog

1-Bit Bonsai: The first commercially viable 1-bit LLMs

Prism ML just released 1-Bit Bonsai models — these are large language models that use 1-bit weights (think extreme compression) while staying commercially viable. The practical upshot: you can run bigger models on smaller hardware. Early days, but this could be a real unlock for edge deployment. Show HN discussion.
Hacker News

Falcon Perception: TII's new multimodal model for enterprise documents

The Technology Innovation Institute released Falcon Perception, a compact multimodal model designed specifically for understanding enterprise documents. If you're building anything that needs to parse invoices, contracts, or other business docs, this is worth checking out.
Hugging Face Blog

⚠️ Also Watching

Microsoft quietly updates Copilot terms: "for entertainment purposes only"

Microsoft's Copilot terms of use now include language saying it's "for entertainment purposes only." That's... quite a hedge for a tool marketed as a productivity assistant. HN is having a field day with this one.
Hacker News

Today’s Sources