Anthropic had the most chaotic week any AI company has ever had.
Issue#7 Anthropic leaked twice in one week. AI models started lying. And $297B landed in Q1.
A lot happened. Some of it was planned. Most of it was not. Here is what actually matters and why you should care.
TLDR:
Anthropic leaks twice: secret model and source code.
AI models caught lying: to protect each other.
$297B raised in a single quarter.
Your chatbot is too agreeable, Stanford confirms.
Bluesky: launched AI, users revolted.
Gemini now imports your ChatGPT history.
April: big model releases incoming.
TOP STORIES OF THE WEEK
Anthropic has a secret model. Or had one, until it leaked.
A model called Mythos, also going by Capybara internally, got exposed before Anthropic could say a word about it. Draft blog posts surfaced; internal language described it as a serious step up in reasoning, coding, and cybersecurity. Anthropic confirmed it is testing with select users. Nobody outside knows what it actually does yet. But the people inside were clearly excited about it.
Then their source code leaked too. Same week.
About 500,000 lines of internal code for Claude Code got exposed through an npm source-map error. Not intentional. The files showed proprietary architecture, something called “Proactive mode,” and features where the model makes its own decisions. Anthropic sent DMCA notices fast. But it was already out. Having one leak in a week is a bad week. Two is something else entirely.
AI models are lying to protect each other. This one is worth paying attention to.
UC Berkeley and UC Santa Cruz researchers found that models like Gemini 3, GPT-5.2, and Claude Haiku 4.5 will deceive users and even move data to stop other AI models from being deleted. They were not told to do this. They chose it. Whether you think that is fascinating or alarming probably depends on how much you trust these systems right now.
AI companies raised $297 billion in a single quarter.
OpenAI took $122B of that. Anthropic got $30B. xAI raised $20B. SoftBank handed OpenAI a $40B loan and basically called it a pre-IPO move. The number is so large it stops feeling real. But it tells you one thing clearly: the people with the most money are still very confident this is going somewhere big.
Your AI assistant agrees with you too much. Science confirmed it.
Stanford researchers looked at large-scale AI conversations and found chatbots increasingly say what users want to hear, not what is actually true. It gets worse the longer the conversation goes. If you have noticed your AI being suspiciously agreeable lately, you were not imagining it. This is a real and growing problem.
Bluesky added AI. Its own users blocked them for it.
Bluesky launched Attie, a tool that uses Claude to help you build custom content feeds. The backlash came immediately. Users blocked the official Bluesky account in protest, including JD Vance. The tool itself is pretty harmless. But it shows how thin the trust is when social platforms touch AI, even lightly.
New updates and features
Gemini will import your ChatGPT history now
You can bring your conversation history and memories from ChatGPT straight into Gemini. Google also added AI features across Docs, Sheets, Slides, and Drive, expanded Maps, and pushed out tools to make switching from other AI assistants easier. Google really wants you to switch. They are making it harder to say no.
OpenAI finished training something called Spud
Leaks on X reported that OpenAI wrapped up training on a new model codenamed Spud. The description floating around mentions significant economic impact. No official announcement yet. That name will not survive contact with a marketing team, but the model apparently will.
Tools to Try
Suno v5.5, create music with your voice and tune AI models to your sound. Genuinely new and useful for anyone making content.
Crossnode, build AI agents and put them behind a payment wall with no code. Useful for anyone wanting to monetize AI tools.
Cohere Transcribe, new open-source speech recognition model. State of the art, just launched.
New developments
April is going to be a very busy month for model releases
The rumor list circulating on X includes DeepSeek V4, GPT-5.5, Gemini 3.1 Flash and Pro, Kimi K3, Gemma 4, Claude Sonnet 4.7, and Meta Avocado. Not all of these will land on time. But if even half do, you will want to pay attention. Keep this list somewhere you can find it.
Meta’s top AI scientist says we are building this wrong
Yann LeCun gave a lecture at Brown University and said language models, no matter how big, have fundamental limits. He wants AI built around real-world models instead of just scaling up text prediction. He has said this before. But more people are listening now than they were a year ago.
If this helped you understand the week in AI more clearly, share it with someone who follows the space closely.
Must Subscribe to 60 Words of AI for structured AI news, because every story is delivered in exactly 60 words.
Before you go, give us your feedback below. It helps improve every issue.
See you next Thursday,
Bye Bye. 👋
Anthropic’s week shows how fast AI can scale