Claude's Leaked AI Modes + Capybara Just Dropped

Claude's leaked AI modes and Capybara. So Anthropic didn't mean to show you any of this. Two leaks in five days. Their internal roadmap is public.

Their unreleased modes are public. Their next model, the one that's supposed to be bigger than anything they've ever built is now public, and it all happened by accident. So let's go through everything. Every unreleased mode from Claude, the new model tier, what it all means for your business right now, because there's a lot here, and most of it hasn't even been covered properly yet.

Number one, you've got Kairos, the mode that never turns off. So let's start with the biggest one. Every AI tool right now works the same way. You open it, you type it, it responds, you close it, it waits.

Kairos ends that completely. So Kairos is a persistent background mode where Claude watches your working environment, writes daily observation logs, and can act on things it notices without being prompted. There's a 15 second blocking budget built in, so it won't interrupt your current flow, and it receives periodic tick prompts where it decides whether to act or stay quiet. So Claude isn't waiting for a question, he's paying attention, watching, deciding what matters.

You're on calls all day. Meanwhile, Kairos is working in the background, noticing something that needs attention, surfacing it the moment you're back without being asked. The name Kairos actually comes from the ancient Greek concept, meaning at the right time. So whilst current AI tools are largely reactive, Kairos allows Claude Code to operate as an always-on background agent.

Not live yet, but it's behind feature flags. So it's built, it exists, it's actually coming on the roadmap for Claude, and it was all revealed inside the Claude Code leak that just dropped. Autodream. So this is where your AI gets smarter whilst you sleep.

Kairos works whilst you're awake. Autodream works whilst you're not. So every night when you're done for the day, Autodream runs a background process through everything Claude learned from that session. The Autodream logic merges disparate observations, removes logical contradictions, and converts its vague insights into absolute facts.

So contradictions get resolved. Stale information gets cleaned out. Vague notes get sharpened into clear facts, which means that you come back in the morning to a Claude that has quietly organized from the day before. Sharper, cleaner, more accurate.

The businesses that grasp this early, that understand that Claude is going to compound in value the more they use it, are going to be miles ahead of people still treating it like a search bar. Then you've also got a ultra plan. So this was also in the Claude Code leak. Pretty exciting stuff.

So what this means is that you've got 30 minutes of deep thinking on demand inside Claude Code whenever you need it. So right now, if you want Claude to do serious strategic planning, you have to walk it through step by step. It's you doing most of the work, right? Ultra plan actually changes that.

So ultra plan offloads complex planning to a remote cloud container running Opus 4.6, which gives you up to 30 minutes to think, and lets you approve the result from browser-based interfaces. So you can hand Claude like a really hard planning problem. It goes away for 30 minutes. It comes back with a structured output and a solo operator building a content strategy, a product roadmap, or a client pitch, for example, instead of like manually prompting through every step, you describe the goal.

It goes off as a little thing for 30 minutes and comes back with a full plan ready to review. That's coming again. It's not live yet. These features have been leaked inside the latest push from Claude Code, and it should be coming in the future.

So the infrastructure is built yet, but it's coming. Multi-agent coordinator is the next feature, and this is a team of Claude's working in parallel. So Claude can already spin up sub-agents. We all know that, but what was in the leaked source is far more developed than what's currently available.

So coordinator mode is one main agent that assigns tasks to multiple workers, and workers execute in parallel and then report back. Workers request permission from the leader via a mailbox system when implementing anything sensitive. So one Claude can be running the show, and then you've got multiple Claudes doing the work simultaneously. How would that work in practice?

For example, let's say you've got a business owner. They're managing multiple client accounts. They could have different agents working each account at the same time. Research on one, copy on another, outreach on a third, all running in parallel, all reporting back to a central coordinator.

Running Claude through one task at a time after this ships is like having a full team and making them take turns, right? Then you've got voice mode as well. So talk to Claude instead of type. So you can, basically what this is gonna allow you to do is talk instead of typing directly into your keyboard, which makes it way faster and easier to use.

So this one is very simple, but it matters more than most people realize. Voice mode is in the feature flags, which means push to talk, real-time voice chat directly with Claude code. For anyone who finds like the terminal intimidating, for example, this changes the whole experience. You just talk, Claude listens, it works, coming, not confirmed yet again, but timeline is coming and it's built, right?

So again, you'll be able to talk directly with Claude code and you'll have this voice mode interacting, which is pretty cool. Next, you got undercover mode and this is a controversial one. So this is the one that made a lot of people stop because inside the leaked sources is a file about 90 lines that instructs Claude to completely hide the fact that it's AI when contributing to public repositories. So the instructions are explicit.

Claude must not include the phase or phrase Claude code or any mention that it's an AI in commit messages or PR descriptions. No attribution lines, nothing. There's no way to force it off. You can force undercover mode on, but you cannot turn it off.

In external builds, the entire function is eliminated. It's a one-way door. The internet had opinions on this one. So hiding internal code names, reasonable.

Having the AI actively present as human and public contributions, that's a different conversation. But the business angle is clear. Enterprise clients working on public facing projects are going to want this. The ability to run AI quietly with no footprint is already built in as a feature.

It's coming as a selling point soon. Now, before I get to Capybara, if you wanna be ready for every single one of these modes, the moment they go live, the AI Profit Boarding is where we make that happen. We've got 2,700 business owners building with Claude right now. The moment Kairos drops, Ultraplan goes live, voice mode ships, Capybara, et cetera, we'll have step-by-step tutorials showing you exactly how to set up each one for your business.

We also got four live coaching calls per week, a 30-day Claude code roadmap built for business owners, daily tutorials, a prompt library, and a member map so you can connect with other Claude users near you and get help 24-7. Link in the description or go to the AIProfitBoarding.com to get access. Now let's talk about one of the most exciting options, which is Capybara. That is coming soon.

And that is a model above Opus on the tier. Now, this one is from a separate source. It's from a separate leaked source code. So there were two different options, and this one came five days earlier from a completely different leak, a misconfiguration in Anthropix content management system, which exposed nearly 3,000 internal files, including a draft blog post that was never supposed to go public just yet.

And what was in that blog post is significant. Right now, Anthropix has three model tiers. So you got Haiku, small and fast. You got Sonic, the middle ground.

You got Opus, the most powerful. And Capybara adds a fourth tier above Opus, bigger, more capable, more expensive. The leaked draft describes Capybara as a new name for a new tier of model, larger and more intelligent than their current Opus models, which were, until now, already pretty powerful. So compared to Claude Opus 4.6, Capybara reportedly gets dramatically higher scores on software coding, academic reasoning, and cybersecurity tasks.

Anthropix confirmed it. A spokesperson actually said they're developing a general purpose model with meaningful advances in reasoning, in coding, in cybersecurity. The leaked draft called it, by far, the most powerful AI model we've ever developed. Anthropix described Capybara as a step change in overall capability.

It is currently being tested with a small group of early access customers. So it's real. It's confirmed. It's being tested right now.

But here's the thing that nobody's talking about. Internal comments from the code leak reveal that Capybara's eighth iteration carries a false claims rate of 29 to 30%, notably higher than the 16.7% recorded in its fourth version. The model is more powerful, but it's also making more errors according to these leaked benchmarks. You have to take everything with a pinch of salt and try and filter through the information, but this is what's out there so far, and I'm giving you that information.

So Anthropix is clearly still working through this. There's also an assertiveness counterweight built in, a mechanism to stop the model from being too aggressive when it rewrites things. That tells you Capybara isn't imminent. They're still actively iterating.

But it's real. It's in testing. And when it ships, the gap between Claude and everything else gets wider. They're already leading.

They're already creating the best APIs in the world. So let's talk about the bigger picture. Let's step back together and look at everything. You've got Kairos, Claude that works without being asked, Autodream, the Claude that gets smarter overnight, Ultraplan, which is 30 minutes of autonomous deep planning, multi-agent coordinator, a team of Claudes running in parallel and voice mode, where you can talk instead of type.

Undercover mode as well, which is a stealth AI for enterprises. And above all of it, you've got Capybara, a new model tier that jumps above Opus entirely. None of this is on a slide deck. It's a compiled list of code behind feature flags that was leaked mistakenly.

They're releasing a new feature every two weeks right now. So who knows what's going to come next? And bear in mind, the public version of Claude is a fraction of what's built. And what's built is about to ship fast.

And here's what I want you to understand right now. The businesses winning with AI right now aren't winning because they found a better prompt. They're winning because they understand the tool very deeply. They've built real workflows.

They've put in the time when Kairos drops, when Capybara goes live, there are going to be two groups of people. People have been using Claude seriously and understand immediately how to use these new capabilities and people are still stuck figuring out the basics. The gap between those two groups is going to be absolutely enormous and it's going to compound fast. Now, if you want to be in the first group, the people winning with Claude, the people who understand how to use these features as soon as they come out, then come into the AI Profitable in community.

There's 2,700 business owners in there building with Claude right now. A lot of them are already running Claude code for client work, content, lead generation and operations. When Kairos ships, when Capybara lands, when Ultraplan goes live, we'll have the tutorials, the coaching calls and the roadmap to show you exactly how to use it in your business from day one. You also get four coaching calls every week, daily step-by-step tutorials, full prompt libraries and a member map so you can connect with Claude users near you 24-7.

And there's always someone online, right? Link in the comments and description or go to the airprofitable.com to get it out right now. The roadmap is out. The features are built.

The only question is whether you're ready to ship with this and whether you're ready when they do ship. I'll see you on the next one.