Claude Max Subscription vs API Usage

[hydrohelix]

I read all the documentation for 5.0 and one of the things that was never mentioned were Claude Max subscriptions. Rather, there are documents pointing to the API usage and costs.

I am assuming this is because of Anthropic's stance on 3rd party harnesses. Is this correct?

Are there instances where parts of PAI can use Claude Code with the built-in max authentication and not require an API key? Such as "claude -p" commands?

Thanks!

[jdrolls]

I use it with the Claude Max subscription. Given that it's an add-on to claude code, at least right now, I don't think there are any issues. With V5 also having telegram integration, and it turning more into a "Life OS" I don't know if we are looking at a change to how it is perceived by Anthropic, I've used Max on it for some time and have never had any issues. It's a little different than services that just use Anthropic's Auth but aren't actually using Claude Code.

The AI scene seems to change every week though so this may be completely wrong or outdated in a month.

[GuillaumePEL]

Hello,

My question might seem silly, but I still haven't found the answer...
Is Claude MAX required? Would a free account be enough? Or do I need a PRO subscription?
Is API billing supported?

[Drizzt321]

I don't see why API billing wouldn't work, since that's built into Claude Code as a base capability. You'd run out of usage very fast on a free account, and Pro might be limiting depending on the model (sonnet vs opus) and your actual use of the system.

So in short, I'd recommend at least the Pro ($20/mo) plan, but you can probably use the Free just to get a tiny bit of experience and see if things might work for you. Personally I'm on Max, and if you're a heavy user I'd recommend that. But it's not necessarily a hard requirement.

[stephtara]

I started out with PRO but very quickly it became too limiting. I do think that if the system is running smoothly and the person using it is not a heavy user, it's possible to make do with PRO (it's my hope). But for me, specially when I moved to v5.0 and the amount of issues I bumped into grew (leading me to a lot of "fixing the system") it was just too frustrating on PRO.

[stephtara]

@GuillaumePEL you need at least PRO to have Claude Code if I'm not mistaken.

[yapierre]

I am using using Max ($100) for first month, because it need a lot at the beginning to setup PAI, Projects and to understand us (TELOS), then only use $20/month after that

[clmitchell808]

I was having persistent problems with Morgan, my personal assistant, failing to follow instructions- or simply forgetting things. I was on the Max subscription and considered moving to the API billing. In the end, I decided to convert the PAI system from Claude Code to Open Code using DeepSeek V4 Pro. I could not be happier. I have asked Morgan to provide an OpenCode translation of the Claude Code PAI 5.0 installation and will upload it as soon as Morgan has it available.

[Raazgar]

@clmitchell808 I was having the same problem with MAIA forgetting things across sessions. Turns out it was likely this: #1192 (she fixed it but still need to test in the next few days if its really gone).

[jwinchester]

I occasionally get close to maxing weekly use on the Max subscription. To economize I created a codex-pai spinoff that lives in a sandbox and can check-out and fork a project from my PAI, write some changes and request merge. Main PAI can then accept the merge pretty seamlessly. Codex seems a lot cheaper in general, especially for coding work. I have not run across any quality or intelligence concerns.

[jdrolls]

Yeah I have a Pi spinoff that I use when I want to play with smaller models and save my usage - I am alwyas hitting that weekly limit. (did opencode for a while, still use it occasionally, but it seems like Pi is a little easier to customize.) Now that other opensource models are getting more usable I have found that for my non-planning or quick play projects I can use a very inexpensive model to do the actual coding. My favorite one over the last couple weeks or so is Minimax M3 - I don't see a ton of people talking about it, but that changes as new models roll out. I run /Evals to bench new models as they come out then use them agains the style of programmign that I do and compare them vs their cost. The nice thing about Pi (or Opencode) is that once you have their version of hooks set up PAI can be used in any harness.

[richardchen874-sys]

@jwinchester This is a very practical architecture. Splitting the main PAI from a cheaper sandboxed coding worker makes a lot of sense, especially if the Max subscription gets close to weekly limits.

For this kind of setup, I’d probably measure the worker by:

• cost per accepted merge
• retry cost when the worker produces bad changes
• how often the main PAI needs to intervene
• latency from task request to merge request
• whether the sandbox protects the main PAI context
• which tasks are safe for cheaper coding models
• which tasks still need stronger reasoning

The key idea is that the expensive assistant should not have to do every coding task directly. It can supervise, review, and merge while cheaper workers handle implementation in isolated sandboxes.

I’m building a China AI API model hub for global builders — one API to access the models independent developers and AI teams need.

If you ever want to compare another lower-cost coding worker setup against your Codex PAI spinoff, I’d be happy to share a small free test credit and benchmark cost, latency, streaming stability, and accepted-merge quality on a real workflow.

[jaylarr176]

@clmitchell808 How has your time with Deepseek been? I am kinda tired of my DA getting mixed up each time Anthropic does something on the back end :(

[richardchen874-sys]

That’s a very interesting migration path. Moving a PAI setup from Claude Code to OpenCode + DeepSeek V4 Pro makes a lot of sense if the main pain is long-running assistant behavior, cost control, and session reliability.

For personal AI infrastructure, I think the key metric is not just “which model is smarter,” but:

• cost per useful session
• whether instructions stay stable across runs
• how much context needs to be replayed
• retry cost when the assistant forgets or drifts
• latency during multi-step tool use
• whether the workflow can separate memory, planning, and execution
• whether cheaper models can handle routine execution while stronger models handle harder reasoning

DeepSeek V4 Pro can be a strong direction here, especially when paired with a workflow like OpenCode where you can control the surrounding system more directly than a fixed subscription environment.

I’m building a China AI API model hub for global builders — one API to access the models independent developers, AI SaaS teams, and automation teams need.

If you keep testing the OpenCode + DeepSeek version of PAI, I’d be happy to share a small free test credit so you can benchmark cost, latency, streaming stability, and session reliability on your real Morgan workflow.

[danielmiessler]

Hey @hydrohelix, thanks for raising this, and sorry it sat for a while.

We're changing how LifeOS ships. Instead of cloning a full ~/.claude directory and running it as a complete system, LifeOS is becoming a skill you install through an agentic installer. The installer hands integration to your own AI, which reads your actual machine (your OS, your paths, your harness) and wires the hooks and system prompt in where they belong.

That's aimed right at what you hit here. The old "one directory, one layout, hope it matches your setup" approach is exactly what broke for so many people, and the new model should handle it far better because your AI does the integration per machine instead of us guessing.

So we're closing this in prep for that release. If it still bites you once the skill-based version is out, reopen or file a fresh one and we'll jump on it. Appreciate you taking the time.