Comment by casparvitch
1 day ago
I guess one issue is that you pay $200/month whether you use it or not. Potentially this could be better for Anthropic. What was not necessarily foreseeable (ok maybe it was) back when that started was that users have invented all kinds of ways to supervise their agents to be as efficient as possible. If they control the client, you can't do that.
I can easily get Claude Code to run for 8-10 hours unsupervised without stopping with sub-agents entirely within Claude Code.
I think it is more likely that if you stick with Claude Code, then you are more likely to stick with Opus/Sonnet, whereas if you use a third party CLI you might be more likely to mix and match or switch away entirely. It's in their interest to get you invested in their tooling.
> if you use a third party CLI you might be more likely to mix and match or switch away entirely.
I really like doing this, be it with OpenCode or Copilot or Cline/RooCode/KiloCode: I do have a Cerebras Code subscription (50 USD a month for a lot of tokens but only an okayish model) whereas the rest I use by paying per-token.
Monthly spend ends up being somewhere between 100-150 USD total, obviously depending on what I do and the proportion of simple vs complex tasks.
If Sonnet isn’t great for a given task, I can go for GPT-5 or Gemini 3.
I don't do this much because I really like Opus 4.5, and so far I haven't hit the limits on the $200 subscription much, but I do have some projects where I might need far higher limits.
As a matter of principle, I really would like the flexibility though, as while I love Opus now, who knows which model I will prefer next month.
I've yet to come up with a workflow where I would want Claude to do this much work... unless I had an extremely detailed spec defined for it. How do you ensure it doesn't go off the rails?
You pretty much just said it. Define an extremely detailed spec. I have one that's five .md files to iteratively churn through. It had to be split into that many files since i don't want to break context length limits on the AI.
On the flip side I started using Claude with other LLMs (openai) because my Pro sub gets maxed out quickly and I want a cheaper alternative to finish a project.
I just use claude code proxy or litellm and set the ANTHROPIC_BASE_URL to my proxy and chose another LLM.
That would seem to be a very good reason for them to make Claude Code good enough that people would prefer doing that over the inverse...
But also, they're a bit schizophrenic about what they want Claude Code to be, given you can stream JSON to/from Claude Code to use it as a headless backend including with your subscriptions.
Multi model is the way of the future though as much as I like and prefer Anthropic.
> I guess one issue is that you pay $200/month whether you use it or not.
I can easily churn through $100 in an 8 hour work day with API billing. $200/month seems like an incredibly good deal, even if they apply some throttling.
Why is supervising one's agents to be as efficient as possible a problem for Anthropic?
When people say efficient here, they mean cost efficient, extracting as much work per dollar from Anthropic as possible. This is the opposite of Anthropic’s view of efficiency, which would be providing the minimal amount of service for the most amount of money.
More inference = more cost (to Anthropic)
What kind of ways to supervise?
ralph-loop & co