Open Claw with Opus 4.6 - Venture Subsidized AI

We had a moment in time where we had access to really amazing software at an incredible price. The Open Claw and Opus combo was just amazing. It was using Open Claw with Anthropic’s Opus 4.6. This was with the Anthropic Max plan at $90 a month.

It was against Anthropic’s TOS but it was commonly understood that using an OAuth token worked with Open Claw. Some individuals got banned (maybe they were hitting it super hard?).

But then on Easter Sunday Anthropic cut everyone off. They gave everyone a $100 credit to continue with their pay for usage model. Plus when you buy credits in advance you get them for 30% off.

This reminds me of when Uber was first becoming ubiquitous. Each ride was subsidized by venture capital. Blitz scaling was the priority - shitty unit economics of each ride be damned. It was easy to get used to that convenience at that price but of course it wasn’t sustainable.

Now grab an Uber in Manhattan at rush hour - the prices can be shockingly high and especially compared to the good ole days.

Someone posted that paying for Opus 4.6 with the same usage as the $90 a month plan was costing over $50 a day.

The cost of AI is real - data centers, energy, GPUs, training models is currently estimated to be in the hundreds of millions of dollars. It’s not impossible we could see training a new model costing $1B dollars. That cost can’t keep being eaten up by and deferred.

So what am I using now? I’m now playing with Hermes Agent with a mix of local and cloud open source models. GLM 5.1 was released just after the Easter day Anthropic massacre. It’s very usable and effective. But nowhere near how good Opus is.

Using iMessage to Edit Code

Voice to Text to Deployed Code

Let’s Unpack

Ok, there’s a lot to unpack there. First, yes I text my agent. And 2nd, yes it’s very fast at writing code.

I had just pushed the first blog post (all coding by AI) and was looking at my blog on my phone for the first time. I quickly realized it wasn’t mobile friendly. I was able to show this to someone and voice to text the fix which appeared almost instantly. She had experience managing developers and developing for mobile and it was just so obvious how this is a huge paradigm shift in software development.

How it used to work

Those of us that have wrestled with CSS, JavaScript, and HTML to get pages to layout just so know the drill. So much trial and error. Cmd-S, Cmd-Tab to the browser window, F5. Rinse and repeat ten times.

And those of us that developed for iOS/Android know an even deeper pain. At one point my mental model developing mobile UI was “best case scenario x 10”.

Fast forward — those days are gone, I do believe.

Vibecoding is a Slur

I’ve become a fan of Peter Steinberger and appreciate his thoughts on leveraging AI for development. There’s a good middle ground between AI-generated code slop and meticulously hand-written code. You have to understand how AI works to really get the value out of it. It’s both insanely capable and limited at the same time. It takes time and practice to get better at using Codex, Claude Code, Cursor, or even OpenClaw over text to make changes.

Make it Easy for AI to Work With Your Codebase

• Code structure & documentation — It’s important to have a clear, logical architecture and structure to your codebase. Essential documentation needs to be pulled out of Confluence and put into .md files in the repo.
• Context efficiency — An agent can’t load a whole complex codebase into context and always starts fresh. Give them what they need to know — but nothing more. With small codebases it doesn’t matter, but this is essential with a large codebase.
• Be specific — Be as specific as possible on expectations, as well as your coding and testing conventions.
• Have honest convos — Agents always try to please, but it’s better to give high-level problems and ask how to best solve them. Make sure it understands the broad strokes clearly to prevent cleanup later.

Thoughts Turn into Code Fast

I remember when I was early in my career thinking about complex coding challenges in the shower. I would then almost have to write down my thoughts for when I got back to my laptop. Now I can just send a text message, which can turn into a ticket to develop later or code changes that get pushed immediately. I can also talk back and forth about different approaches with the agent before development. Or I can just update a .md file in the repo to tackle things later.

Coding via voice-to-text over iMessage is pretty rad. Your phone stops being a consumption device and starts being a development environment.

Do I Have To Buy a Mac Mini to Run Open Claw?

I’ve started plaing with Open Claw and having a ball. Actually inspired me to start blogging again after taking looooooooong break. For those unfamiliar it’s an open source personal always on agent that can live in the cloud….but it seems most of us are running on local computers. No more giving so much data to Open AI or Anthropic.

The Open Claw forks are exloding - most notable Nemo Claw by NVIDIA.

Open Claw runs great on a mac!

You can run OpenClaw on whatever computer you already have — an old laptop works fine. Lots of folks even using raspberry pis. With more horsepower come more options to run local LLMs. More on that later.

Why Mac Mini

I looked at small PCs for comparison — and it’s a mess. I genuinely didn’t know what the best equivalent would be. If I wanted to train a model, sure — I’d be looking at NVIDIA GPUs. But that’s not what this is. I wanted simple, fast, and easy. Mac Mini.

But any computer will work - mac, windows, linux.

It’s All About Memory

Apple Silicon’s unified memory means you don’t have to think about RAM, GPU, VRAM, whether your model fits in memory, or which configuration matters. On a Mac Mini there’s no separation. Memory management is nothing you need to think about.

The Hunt

I naively thought I could walk into an Apple Store and pick one up. That’s not a thing.

I ordered one in early March. Delivery kept slipping — March 13-20 became April 13-20. I started watching different configurations and one day a next-day shipping option appeared. I jumped on it.

What I’m running: M4 Pro, 12-core CPU, 24GB unified memory, 512GB SSD.

There’s a run on these right now. No Mac Mini configuration is currently available until April. The M4 Pro models aren’t until June. Apple seems so far behind on AI - Siri is a joke now. But maybe apple will sell a ton of these. I bet they were confused at first on why so many were being sold.

Where It Lives

Right now it’s on a windowsill catching some Austin sun, running over wireless while I get things dialed in. Once it’s stable it goes onto wired ethernet, plugged into the router, and disappears. My own personal AI that’s just a text away.

More on what it’s actually running…