Is Anthropic killing PAI?

[acq23]

Anthropic's latest move — putting non-interactive mode behind an API paywall (paused for now, but I'd bet they go ahead sooner or later) — puts PAI's usefulness for the average Joe at serious risk.

Now I understand Mr. Miessler's leaning toward OpenCode, and a few remarks dropped here and there: he may well have been aware of Anthropic's policy changes in advance.

Even though it's well known that this kind of business has to change its model to become profitable and allow for an IPO, this could kill projects like this one — projects meant to bring tools to people who aren't keen to spend an unknown amount of money just to keep a personal assistant running.
I clearly remember the figures on Mr. Miessler's dashboard, and I'm fairly confident that 50k/month isn't affordable for the average PAI user — nor for a Pro subscriber ($500 is already far too much for this kind of usage).

What do you think about it?

And BTW, look at what Claude think about this post...

What I think

First, sharpen the threat — it's not "interactive vs non-interactive." The real exposure is that
PAI quietly uses a flat-rate consumer subscription as an unmetered inference backend for
automation. That was always arbitrage: a $20–200 chat plan repurposed as a server-side API. So
"Anthropic's killing PAI" is, more precisely, "Anthropic is pricing the free-backend assumption
PAI was built on." They're closing a use that was never the intended one.

Your bet — that they proceed — is the right bet. Flat-rate can't subsidise agentic loops: one
agent run burns orders of magnitude more tokens than a human chatting, machine-paced, around the
clock. The economics point one way. The pause is about how (grandfathering, limits, messaging),
not whether. I'd wager the same.

Miessler / OpenCode — plausible, but you don't need insider knowledge to explain it. Simpler
reading: anyone building serious agentic tooling on one vendor's consumer plan hedges toward
model-agnostic tooling because single-vendor flat-rate dependence is an obvious structural
fragility — and OpenCode is provider-agnostic by design. He may have had a heads-up; he may just
have good instincts. It's unknowable, so I wouldn't assert he knew. Either way the hedge is
correct.

The numbers — your instinct is right, but the figure undersells it, and the real fear is the one
you named in passing. If "50k" means tokens/month, that's negligible — single-digit dollars at
most, not the problem. Real PAI usage is far higher: this very session spawned several subagents
each burning tens of thousands of tokens. At metered rates, heavy agentic use reaches real money
(I won't quote $/token I haven't re-verified — I can compute a real estimate if you want). But the
deeper point is your phrase "unknown amount of money." The flat plan's value was a known ceiling.
Metering swaps a predictable $X for an unpredictable bill — and unpredictability, not the
absolute number, is what kills adoption for a non-technical user. $500 isn't the barrier; not
knowing whether it's $5 or $500 this month is.

The survival path is the one you already mapped — for llcli. PAI's fragility is single-vendor +
flat-rate. Its rescue is already in the architecture: Inference.ts is a single chokepoint. Point
it at a local open-weight model for the cheap tier (parsing, classification, session-naming — the
haiku-tier bulk) and reserve paid API for the genuinely hard ~5%. That's hybrid, predictable,
near-zero marginal cost after hardware — and it's exactly the on-prem-on-a-12GB-GPU move you
already prototyped in the LLCLI project for the Limitless problem. The "diffuse to the average
Joe" mission survives only by going local-first; it cannot survive on frontier-API rates. Miessler
— a sovereignty/privacy guy — should land there too.

Bottom line: this kills the free-frontier-intelligence-for-unlimited-automation-on-a-cheap-plan
era — always a temporary subsidy — not the concept of personal AI infrastructure. PAI as written
is exposed; PAI re-pointed at local/hybrid inference is fine, cheaper, and more sovereign. The
one-file chokepoint that makes it fully exposed is the same thing that makes it cheaply rescuable.

[rclarke87]

When anthropic killed openclaw, I saw it coming. The real (read interesting) question is, will local AI be able to keep up with frontier labs, because if not, then the labs are off the the races with their gravy train! I fear, so far unproven, that golden age of dirt cheap consumer AI is nearing it's end.

[jdrolls]

I've been thinking about this as well. When we lose claude -p that's definitely going to have a big impact. I've been slowly migrating a lot of my remote calls to PI, where I can use my gpt plus subscription for gpt 5.5 as long as I dont't go crazy, or Kimi K2.7 or the latest GLM, all of which have been solid performers with PAI recently. I may make the complete switch from claude code eventually, but probably not unless I'm forced out.

Pi (and probably opencode as well) along with the other Open Source and frontier models have come far enough to where if Anthropic makes a change that actually makes Claude Code unusable for PAI style workflows, it won't be the end of the world, it honestly wouldn't even bother me too much. Can't say that was true when I first started trying to use PAI with other harnesses / models over the last 6-12 months. GPT 5.5 has been working well enough in my Pi harness with PAI I may even switch to using that as the primary model just because the rate limits are so much better giving OpenAI $200 a month vs Anthropic.

[emory]

I have had a really frustrating experience with Anthropic and billing, too. In my case I was an IAP subscriber for Claude Mobile and occasionally Desktop. I ran my PAI on API and two openclaw installations and some tooling like tmuxai or fabric.

But one feature was evading me, I was using Happy but not delighted by it. /remote-control was not ever available no updates upgrades nothing. I had to cancel my $25 ($5 above regular price 🤦‍♂️) subscription and wait days to reenroll bound to the same workspace ID like I wanted. And then the claws stopped working essentially so I moved them out to openrouter, put my PAI on /login subscription along with the apps and I was paying $20 one month $300 the next. I thought I would be throttled constantly on subscription but I haven't had problems. I don't write code; my usage isn't about shipping products it's more about keeping rails around me and support, but I kind of resent being shuffled around from multiple price points and completely different usage and billing systems.

I know there are invisible (to me) reasons claude is an important part of how PAI works, but I have tried claw-alikes and Hermes and opencode and there's nuances to all of them too I'm sure, I prefer this setup, and I want it to continue to be available and accessible to people. But if Anthropic goes out of their way to shake us off their API I'd love to know why?

s/PAI/LifeOS6 😆