What Launched

On September 22, Anthropic released Claude Opus 5.5 and cut its price 20%: $4 per million input tokens and $20 per million output, down from $5 and $25. Later the same day, OpenAI released GPT-6 Sol and Luna at half the price of their predecessors.

Both companies say the new models are better. Those claims are theirs, not mine, and I haven't tested them against each other. What I can check is the price, so I did.

What It Means for a Small Business

Take a typical server task: reviewing a config file. Say 20,000 tokens in, 2,000 out. On the new pricing that's about 12 US cents. On the old, about 15.

That's the honest answer for a small server owner. If you pay per token, the price cut saves you pennies. Nobody should choose a tool over three cents. Choose it on whether it gets the job done right.

What I Actually Use It For

Thirty years of running my own servers, and the last months with an AI crew beside me. Real jobs so far: a NodeBB forum installed as a systemd service behind Nginx with SSL. A script that keeps a firewall rule current for an MQTT broker when Starlink won't give me a real public IP. A migration of Google Photos into WordPress with the metadata intact.

None of it was magic. All of it was faster.

Different Tools, Different Strengths

I now use Claude, GPT, Gemini, Muse and GLM. Each has its strengths and weaknesses. The skill is knowing how to use them, and which one is best for you and your needs.

Case in point: an AI once drew the Benfica eagle as a dove. It was Claude. I saw the mistake in seconds. The reviewer never did. The logo was an SVG, and when the reviewer read the numbers, everything said "bird". True, as far as it goes. I had to take a screenshot to show it was a dove, not an eagle.

The lesson I took from it is simple: don't use Claude for logos. GPT and Gemini are better at that. Claude, in my experience, excels at code and writing. And checking the code is not the same as looking at the result.

Same rule as any toolbox. You don't pick the hammer because it's the newest one.

Where It Still Gets Things Wrong

Early on, a different assistant spent hours on a coordinate-parsing bug in my Journey app. Confident fix, same bug back with a new line number, confident fix again. A loop that sounded like progress.

Every model does some version of this. The output sounds sure of itself whether it's right or not. You need someone who knows what right looks like.

cPanel Lets Agents In

On September 17, cPanel declared cPanel AI generally available, including an AI assistant and MCP access, the same protocol I use for my own server. An agent can now work inside a hosting account on a standing grant of access.

I'm not against it. I run MCP myself. But "standing access" is the phrase I'd read twice.

What Comes Next

I plan to run local AI models on my own servers, and I'll report the results in future posts. 😉

My Rules

  • Start read-only: logs, configs, reviews. Let it look before it touches.
  • Backups before anything that writes. Always.
  • Billing and customer data stay out of the agent's reach.
  • Match the model to the job.

The models will keep getting cheaper. The habit that matters costs nothing: know what the machine is allowed to do, and keep a human hand on the parts that can't be undone.

ALL SYSTEMS OPERATIONAL.