Launching the Claude Partner Network

Nobody would use a $1k sub if using the API pricing would only cost $500 for comparative service.

For the record, I'm only explaining what he put forward.

I don't agree with the opinion, mainly for two reasons:

The API cost can be increased in conjunction, hence the ceiling is just as variable

The harness is even more important then the model ime, and Claude Code is getting better every month. Even though the alternatives are getting better too, they're at least currently significantly worse IME - I'd say at least 3-6 months behind (compounded by the model, ofc).

And as a third point, unrelated to the original argument: there is no way anthropic is actually treating the sub as a loss leader. It is not cheap. It's only cheap compared to their API pricing, which they can freely set however they want. Compare their pricing to free models like Kimi k2.5 etc. I sincerely doubt anthropics model costs more to run then theirs, and they're profitable at 30% of the price anthropic charges.

It was cheaper prior to them issuing certificates, then it got expensive.

An hour ago Gemini decided it needed to scan my entire home folder to find the test file I asked it to look into. Sonnet will definitely try to install new dependencies, even though I’m doing SDD and have a clear AGENTS.md.

I’m always baffled at people’s magic results with LLMs. I’m impressed by the new tools, but lots of comments here would suggest my Gemini/Sonnet/Opus are much worse than yours.

As the same age as Linus Torvalds, I'd say that it can be the opposite.

We are so used to "leaky abstractions", that we have just accepted this as another imperfect new tech stack.

Unlike less experienced developers, we know that you have to learn a bit about the underlying layers to use the high level abstraction layer effectively.

What is going on under the hood? What was the sequence of events which caused my inputs to give these outputs / error messages?

Once you learn enough of how the underlying layers work, you'll get far fewer errors because you'll subconciously avoid them. Meanwhile, people with a "I only work at the high-level"-mindset keeps trying to feed the high-level layer different inputs more or less at random.

For LLMs, it's certainly a challenge.

The basic low level LLM architecture is very simple. You can write a naive LLM core inference engine in a few hundred lines of code.

But that is like writing a logic gate simulator and feeding it a huge CPU gate list + many GBs of kernel+rootfs disk images. It doesn't tell you how the thing actually behaves.

So you move up the layers. Often you can't get hard data on how they really work. Instead you rely on empirical and anecdotal data.

But you still form a mental image of what the rough layers are, and what you can expect in their behavior given different inputs.

For LLMs, a critical piece is the context window. It has to be understood and managed to get good results. Make sure it's fed with the right amount of the right data, and you get much better results.

> Nowadays I just paste a test, build, or linter error message into the chat and the clanker knows immediately what to do

That's exactly the right thing to do given the right circumstances.

But if you're doing a big refactoring across a huge code base, you won't get the same good results. You'll need to understand the context window and how your tools/framework feeds it with data for your subagents.

For any proctored standardized testing a person takes, AI should be able to quickly summarize that person’s abilities. This way, instead of people writing their own BS resumes, a trusted test provider can evaluate an individual deeply, solving the problem of having to waste time on coding interviews etc. it will speed up hiring.