Why we run Claude Opus 5.5 and our own Jah 6.0

Here's how the RMDW team decides which AI touches a customer's work. It's one question.

Does it need to stay private?

If the answer is no, we use the best public AI there is. Right now that is Claude Opus 5.5. If the answer is yes, we use our own model, Jah 6.0. Private work stays private.

That's it. That is the whole stack, and it has taken my business to a different level. Let me show you how we got here and what it means for yours.

Opus 5.5 made AI fun again

I worked with Claude Opus 5.5 for a week before I recorded this, and I have to be honest. It is amazing.

It has changed how I shoot videos. It has changed what we hand to clients. It helped me make a real advertisement for RMDW. Not some regular old thing. Real video, real branding, the works. It helped me improve our own tools.

And it is a big part of the client work. Customers come to us asking for websites, because their website is how the world sees their business. They ask for apps that work inside their business. Think CRMs, dashboards, booking, invoices, calculating their data. Opus is right there for that.

Look, y'all heard me talk trash about Anthropic. I still have my problems with them. But they came out with Opus 5.5, and I cannot deny how good it is.

Why we still run our own model

Now. Opus has limits, and not just technical ones.

The day I recorded this, I was working with Claude on a podcast and it pushed back on me about a voice. "Well, I don't know, what is this voice?" Dude. I made the voice. Claude still has its foolishness.

And some work just should not go to someone else's servers. Personal work. Client files. Anything a business would not want floating around.

That is where Jah 6.0 comes in. It is our own model. Whenever the work is private, whenever I need the AI to just do the job without the hand-wringing, Jah is there.

A few facts about Jah 6.0 you can check:

• The weights are public on Hugging Face.
• It powers Kiyomi, our private AI for offices with sensitive files.
• On RMDW Bench v1, 20 real coding tasks where 19 out of 20 is the ceiling, it scored 19 out of 20. That is level with Claude Opus 5 and GPT-5.6.
• It plugs into tools like Claude Code, Codex and Cursor.

A month before this video, I was out here trying to prove whether Jah was better than Anthropic and OpenAI. I don't need to anymore. With Opus 5.5 helping the work and Jah handling the private side, I literally don't need ChatGPT. ChatGPT is the odd man out for me right now.

The model is not where the money is

Here's something I believe. There is no money in AI in the sense of just having a model.

I'm not going to say everything, but something like 95% of the result depends on the harness. The setup around the model. Get a great harness, and a model with the smarts of Opus 4.8 can perform above Fable level. I'm living it.

And business owners do not care about any of this. Most people don't know what Opus is. Most people barely know what Claude is. Small, medium or enterprise, they want the outcome. They want it done.

Then the very next thing they ask: can it be private? Doctors, lawyers, people in education. They want it private.

So that's how we split the work. If a customer wants it done and it's not sensitive, boom, Opus. If it's private, boom, Jah.

And the assistant in my pocket

There's one more piece. What about the AI I just want to talk to? The one that handles small things, checks on stuff for me, or goes out and looks for customers when I don't feel like it?

That's Meta's Muse. I have to give it to Meta. Muse is incredible. If you're technical, you know Muse is basically OpenClaw. But it is so much better than OpenClaw, it's ridiculous. It's what OpenClaw should have been. I pull out my phone, tell it what I need, and it works in the background with its little character and takes care of business.

Now, when you hand it passwords and information, it says it stores them somewhere it can't see. Is that true? I don't know. That is exactly why anything sensitive belongs on the private side.

So my stack is Opus 5.5, Muse and Jah. The Claude plan is 200 dollars a month if you really want to use it, and Muse is free right now.

Don't pick a side. Go hybrid.

This is where I had to change my thinking. You don't want just public AI. You don't want just private AI. You want both.

When you need the best quality and it isn't sensitive, the frontier models win. When something needs to stay with you, your own private AI wins every time.

What this means for you

Sort your work into two piles: public and private. Use the best public AI for the first pile. For the second, you need AI that you control.

That's what Jah 6.0 is for. It is RMDW's own model, it is behind Kiyomi for offices with sensitive files, and the RMDW team can put it to work on the private side of your business while the public side gets the best AI there is.