How much does pi cost you when you’re using 5.5 via API?

I feel like the max CC plans are incredible value.

Why would you use it via the API?

I personally don't use Pi but Zed, but Claude is plugged using ACP on my Pro plan. I would never pay the API prices that's way too expensive.

How are you doing this?

I was under the impression that using Pro or Max from any harness but Claude Code was clamped down upon and likely to result in a ban so haven’t pursued it.

[Answering myself…]

Huh, I guess not.

You can still use harnesses/adapters that use ‘claude -p’ to connect behind the scenes. But the billing is extremely unclear, and it sounds like they use ‘more plan per token’ than just using Claude Code.

The Support page is very unclear, mentioning that you now get an allowance of tokens (and not mentioning the Pro plan at all in that section), and also saying that they’re backing off on plans they previously announced in June without saying what those those plans were, or what the status quo before them was:

https://support.claude.com/en/articles/15036540-use-the-clau...

…and there’s also this GitHub issue showing that tokens are ‘more expensive’ when using `-p`:

https://github.com/anthropics/claude-code/issues/97074

[delayed]

Isn't plan mode just a "make a plan first, don't change anything yet" addendum to your prompt? Haven't tried Pi, but given it's famous extensibility, that should not be too hard to build.

If you really want a “plan mode” you can ask Pi to add it. This is the strength of Pi. It isn’t opinionated far past excellent execution.

Anything extra you can add yourself. You just ask whatever model you’re using to query the documentation included in the pi application and then its away.

My base instructions for all harnesses say something like "when the user asks you to do something that seems like it would be complex, come up with a plan first."

Works.

I switched to Pi and it was great but I had to switch back to Claude Code purely for the subscription. Even with all the token efficiency I built into Pi it just wasn't nearly as good value as an Anthropic subscription. All open models ended up being more expensive so I switched to using an OpenAI subscription, which was better, but the OpenAI subscription is way less generous AND their models are behind. It became hard to justify.

Sorry, what on earth are you doing where the open models are more expensive than the subs?

When I was using anthropic subs, my "equivalent usage" was $6k/month.

On the open models, it's < $100/month.

Same work (actually, more.)

I’m comparing the options available to me: Anthropic subscription vs open models via API. If you have to pay via API then open models are cheaper, if you don’t then an Anthropic/OpenAI is cheaper and far better.

No, my point is I can get the same work done paying < $100/month with the open models.

The large open models (GLM-5.3, Kimi3, etc) are far more expensive than their large closed equivalents (Opus, Sol), because the closed models are far more subsidized when using subscriptions.

If you see no difference at all in between Opus and Deepseek Flash for your work, that is great! Personally, I found the latter to be unusable for the projects I was working on.

Can you describe what you are doing where comparing the two is problematic?

For what it's worth, I use DS flash for statistical analysis, design, scripting, server maintenance. The same I used for Opus.

I switch harnesses regularly depending on the model providers, and I only use skills which are universally usable in any model and harness combination. I don't want to get tied to a particular model or harness, and skills provide me the freedom to do so.

I can include any custom programs I need in a skill, and the SKILL.md file is just a doc of how to use those. It helps me save tokens and ensures determinism with repeated workflows since redoing a task is just calling a skill's program.

I also load only skills that are necessary for a project. I don't want to pollute the context window with unneeded skills.

> Luckily you can write them yourself pretty easily.

According to the article you can ask Pi to write ssuch extensions, so you don't have to, right?