MatIntellect

⚡️ Today Anthropic introduces separate billing for Claude SDK - $100/month per account. I moved in advance, and in the process discovered something most developers will miss until they get the bill

Details of the switch to Claude CLI:
A month ago Anthropic announced: starting June 15, all calls through Claude SDK go on separate billing. Sounded like you just flip a config and forget about it. In reality it turned out to be nothing like that. Many client projects were built entirely on Claude SDK - I had to migrate them first, and only then deal with my own system. The process took way longer than I expected - especially where the SDK was baked deep into the automation layers

What I found during the system audit:
✅ Client projects on Claude SDK - the first and most labor-intensive migration priority
⚡️ claude -p in automation tools is also Claude SDK, and there are way more of these calls inside the system than it seems at first glance
💡 $100/month per account starting today - under real production load it won't last even a few days

The biggest surprise was exactly claude -p. Automation tools call it where you don't expect it and don't even think it's an SDK call. Run grep across all your configs and scripts right now. Makefiles, bash scripts, workflows - that's where claude -p shows up most. Migrating one process takes 10-15 minutes if you know where to look. Find it and move it to plain Claude CLI

Why this is just the beginning:
I expect the screws to keep getting tighter. Even the methods I'm using right now - Anthropic will eventually change or cancel those too. This isn't paranoia, it's a pattern: every few months the terms change and you have to move again

The only way out is to build a system that doesn't depend on a specific API. So my next stop is a multi-LLM architecture where each process calls its own model. Different tasks - different models. I've almost built a system like that, and it's a whole other level of independence from any provider

💡 Deadline: June 15, 2026 - that's today

How the multi-LLM system turned out

The promised system is running. I now have 8 Claude Code agents working every day, and their brains are Sonnet 5 and Opus 5.5 for different roles: where depth is needed the stronger model works, where speed and volume matter a fast one is enough. Speech has its own model entirely: I transcribe long calls on my own computer, a 57-minute conversation becomes text in about 10 minutes, and clients' voices never leave it. The summary and tasks from that text are then written by a language model

The main lesson of the move is built into my personal agent Jes. It has a primary model and an automatic backup: if the first fails or hits a limit, the second carries on by itself. Jes has been running for 134 days and handles 10 scheduled jobs on its own, and no pricing change has stopped it in that time

How I do it for clients

For clients I set the system up on their own server and their own AI subscription, so no provider price change can bring it down overnight. That's how AI assistants run for five people, and how a clinic owner runs 6 agents. The model inside can be replaced without touching memory, rules or the schedule, because those are stored separately. If prices change tomorrow, the client simply switches the model instead of moving in a panic like I did back then

So the first question I ask on any project is not which model is best, but what happens if that model doubles in price or disappears tomorrow. The answer should be simple: switch to another one and keep going

🗒 I've almost built a multi-LLM system where each agent calls its own model for its own task - I'll tell you more soon. If you don't think about architecture right now, later you'll be moving in a panic under new prices and restrictions. Better to do it with a cool head

Share:

No comments yet

Leave a Comment

Fields marked with an asterisk (*) are required