Hacker News Reader

Comment by yojo | Hacker News Reader

Claude Code injects a ton of tools into the system prompt, including their “memory system” that’s like 10k+ tokens. Depending on your task shape, this can easily double your task cost (e.g. a low context-using job that takes many turns, like a monitoring loop).

You should use --disallowed-tools to prune any tools not needed for the task. Note that this is also a perpetual game of whack-a-mole since they’re always adding new tools.

View on HN

> Claude Code injects a ton of tools into the system prompt

It's so unfortunate they don't let you use the subscription with other harnesses anymore - since even if I used OpenCode they'd still get a bunch of useful data from the API calls, meanwhile I could stretch their tier limits way further.

It's so unfortunate people don't realize it's cheaper to write their own well working agent instead of insisting with general purpose bloated ones like Claude.

200$/month is a lot of money on Luna/DS4 flash, like really a lot and the results are much better than clowning on bloated CC.

It's absurd how you have more and more organizations encoding their processes on LLMs and "engineers" (charlatan coders) don't even bother optimizing the tool they use most.

> the results are much better than clowning on bloated CC.

I won't argue with the cost effectiveness, but the results are very much not better. Opus and Fable are in a different league than DS4 Flash. Even GPT Terra, which I really like overall, sometimes gets stuck in weird loops and starts to do stupid stuff once its context window fills up. Whereas I can more or less trust the big models to just Do The Thing™ on the first try.

With that said, you get way more value out of a GPT subscription than you do from Claude, partly because of the ability to use more efficient harnesses.

You're confusing models with the agents.

You can use opus or fable if you please, I'm advocating for writing your own agent instead of used generic bloated ones like CC.

I'm not confusing anything since you can't use custom harnesses with your Claude subscription—you have to use Claude Code. So as far as the Claude sub is concerned, the models and the harness are coupled.

I have yet to spend $200 on DS4 after two months of using it with an entire team.

It's cheaper still, and just as effective, to write code yourself instead of having the LLM do it for you.

Thanks, you make me feel better for spending the (fun!) time writing my first harness in Emacs Lisp with a ‘Emacs UI’ and later writing a command line harness in Common Lisp. Am I more productive with my own harnesses? Not yet, but the second harness I wrote in Common Lisp is promising, and I might write a little book just on this project to encourage people to hack what I wrote and make it their own.

Subsidized Anthropic subscriptions seem to work fine on the oh-my-pi harness, somehow.

Calling per token usage of the US closed source labs has always been funny to me, we have a clear model of what it actually costs to host these models from open models.

Your subscription is not subsidised, it is just closer to the actual cost of the model…

Really? I immediately got a warning that the tokens would come from extra usage, and noped out immediately (so maybe pi lied to me?)

According to ccusage I use about $5k of tokens on my $100/mo Claude Max plan and only hit 5-hr windows where I have to switch tools for a couple hours about once a week

The funny thing is that in some cases using a different harness with the subscription plan could actually be very good for Anthropic: e.g. if I were to use smol with Opus, it could use fewer tokens than CC for the same task. People on subscription plans burning fewer tokens is a good thing for Anthropic.

The only downside for Anthropic that I can see is that hitting your limits more often (while using CC) could make you want to upgrade plans, and a more efficient harness could keep you from doing that. But I can't imagine the cost (to Anthropic) of those inefficient tokens is worth it to them.

The system prompt part is surely cached across all users.

For Anthropic models, yes. But OP was using Claude Code with GPT.

You still pay cache token costs on API calls. Cache cost/token are 90% lower, but you pay it every single turn.

I’m not sure if they let you skip the cache write cost on the first turn. That would imply cross-user caching infrastructure or special casing the default system prompt to give you a discount. Maybe? Away from the computer but you could try a “hello” in a fresh session and see what was billed.

Its inexpensive to reserve KV cache for the first turn and would benefit users, given the first turn already requires costly locating and allocating a model slot for a user.

So yeah, when they banned tgird party harnesses there was a technical and $ case to have.

the system prompt still takes up useful space in the context window and steers the model into unnecessary actions and over-thinking patterns

ty re --disallowed-tools

for coding agents 'shell' is often all you need (just make sure the environment has the necessary tools)

appreciate the tip, kind netizen

Reply on news.ycombinator.com