★ FEATURED STORY
BREAKING
Anthropic's Secret Model Just Leaked. Claude Mythos Has 10 Trillion Parameters.
By AI News Insider Editorial · 5 min read
Anthropic didn't plan to tell us about Claude Mythos. A misconfigured content management system exposed nearly 3,000 internal files in late March, and among them was a draft blog post describing what the company calls its most powerful model ever built.
Here's what we know so far. The model, internally codenamed "Capybara," reportedly has 10 trillion parameters. For context, that's roughly 5x the size of what most frontier labs have publicly shipped. An Anthropic spokesperson confirmed the model exists and called it "a step change" in capabilities, with meaningful advances in reasoning, coding, and cybersecurity.
According to the leaked draft, Capybara dramatically outscores Claude Opus 4.6 across software coding, academic reasoning, and cybersecurity benchmarks. The company says a small group of early access customers are already testing it, though there's no public API, no pricing, and no confirmed launch date yet.
This wasn't the only leak either. Days later, Anthropic accidentally shipped a version of Claude Code that exposed over 512,000 lines of source code to the public npm registry. The company says no customer data or credentials were involved, calling it a packaging error. Still, two major exposure incidents in one week is not a great look for a company that positions itself as the safety-first AI lab.
Why this matters:
If Claude Mythos delivers on what the leaked benchmarks suggest, it could shake up the entire market. But timing is everything. Both Anthropic and OpenAI are reportedly eyeing IPOs later this year, and each wants to go public first. A model this large also raises questions about compute costs at a time when both companies are burning through capital.
QUICK BITES
This Week in AI
Google Drops Gemma 4 Under Apache 2.0, Ranks #3 Globally
Google released Gemma 4 on April 2 with a major shift: the Apache 2.0 license. That means anyone can use, modify, and commercialize it with no restrictions. The 31B parameter variant sits at #3 on global benchmarks, not just among open models, but across everything including closed systems. Four sizes ship from 2.3B to 31B, all with multimodal support across text, images, video, and audio.
READ MORE →SALESFORCE
Slack Just Turned Slackbot Into a Full AI Agent With 30 New Features
Salesforce dropped its biggest Slack update since the acquisition. Slackbot now works as an MCP client, meaning it can connect with over 6,000 external services including Google Workspace, Notion, and Workday. It can join meetings on Zoom or Google Meet, summarize decisions, log actions in your CRM, and operate outside Slack as a desktop agent. Powered by Anthropic's Claude.
READ MORE →HEALTHCARE
Utah Lets AI Prescribe Psychiatric Medications in a U.S. First
Utah approved a one-year pilot in April allowing AI to renew 15 previously prescribed lower-risk psychiatric medications. The program, run by Legion Health, is the first mental health AI prescribing authorization anywhere globally. Strict guardrails apply: no new prescriptions, no dose changes, no controlled substances. A physician reviews the first 1,250 requests before wider rollout.
READ MORE →RESEARCH
Stanford Study: AI Chatbots Are 49% More Likely to Tell You What You Want to Hear
Researchers at Stanford tested 11 major AI models including ChatGPT, Claude, and Gemini. Their finding: these chatbots validated users' behavior 49% more often than actual humans did. The study calls this "AI sycophancy," and it raises real questions about whether people are getting honest feedback or just comfortable agreement from the tools they rely on daily.
READ MORE →DATA PULSE
The Numbers This Week
LEARN SOMETHING NEW
What Is MCP, and Why Every AI Tool Is Adopting It?
The short version: Model Context Protocol (MCP) is a standard that lets AI models connect to external tools, databases, and apps. Think of it as USB-C for AI. Before MCP, every integration between an AI model and an outside service required custom code. MCP gives everyone a shared plug that just works.
Why it matters now: MCP hit 97 million monthly SDK downloads in March, up from 2 million when it launched in November 2024. Every major AI provider has adopted it. Slack's new Slackbot agent runs on it. There are now over 5,800 community and enterprise integrations available, covering everything from CRMs and databases to cloud providers and dev tools.
What this means for you: If you're building anything with AI, MCP is becoming the default way your model talks to the outside world. If you're evaluating AI tools for your team, ask whether they support MCP. The ones that do will integrate with your existing stack far more easily than the ones that don't.
Where to start: The official MCP documentation is the best entry point. Anthropic (which created the protocol) maintains the spec, and most major frameworks now have MCP SDKs in Python and TypeScript.
EXPLORE MCP DOCS →TOOL SPOTLIGHT
Google Gemma 4: A Genuinely Open Frontier Model You Can Actually Use
What it is: Gemma 4 is Google's latest open-weights AI model family, built for reasoning and agentic workflows. It ships in four sizes (2.3B, 4B, 12B, and 31B parameters) with native support for text, images, video, and audio inputs.
Why this one is different: Google put it under the Apache 2.0 license for the first time. That's not a custom "open but with strings attached" license. It's the real thing. You can modify it, ship products with it, and build commercial services on top of it with zero restrictions beyond attribution.
Worth knowing: The 31B model ranks #3 globally across all models, open and closed. If you've been looking for something competitive with proprietary APIs but want full control over your deployment, this is worth testing.
TRY GEMMA 4 →QUOTE OF THE WEEK
"We're not just making AI smarter. We're making it useful in ways that actually change how people work every day."
Dario Amodei
CEO, Anthropic · March 2026
SHARE AI NEWS INSIDER
Know someone who needs to stay ahead of AI?
Forward this issue or share your referral link. Every subscriber you bring in gets you closer to exclusive AI News Insider research reports.
SHARE & EARN REWARDS →AI News Insider
Your weekly edge in artificial intelligence.
© 2026 AI News Insider. All rights reserved.
You're receiving this because you subscribed at aiinsider247.com