Okay, while I don’t plan on waking up most Saturday mornings with a special edition that I just feel like I need to get out there…this is one of those mornings. While I think there’s almost no chance you missed it, if you say, were at a retreat, with no access to Internet for the last 24-hours and you just re-connected to the world, this is what happened yesterday, in one sentence.
Anthropic emailed customers letting them know, that using a regular Claude subscription plan, with OpenClaw, is no more, and it starts today.
A LOT of people were (understandably) shocked, and very quickly news started spreading claiming that Anthropic had banned OpenClaw, which isn’t quite true.
The reality is - you can, and many people will continue to, use Claude models with OpenClaw, it just needs to happen via the API. Boris, the creator of Claude Code was doing damage control pretty much all Friday night clearing up confusion.
Boris did a deep dive into why the decision was made as well, across multiple tweets, here’s one of them:
Thariq, a member of the Claude Code team, also was busy on X last night helping to clear up confusion and do damage control. And yes, Anthropic is issuing a month of free credits.
But, it kinda goes without saying that whoever decided to make this decision, and make it effective within 24-hours, probably didn’t think through it well enough. Personally, I get it - Anthropic was probably starting to lose a fortune on OAuth users on OpenClaw, and yes, it was against their TOS…but they could have likely given people more of a heads up and put together some tutorials on how people can use Claude the way they want them to with OpenClaw.
Instead, waking up this morning, I see tweets from people like Garry, the CEO of Y Combinator that are definitely powerful.
And, OpenClaw Pros like Claire Vo, who recently went on Lenny’s Podcast to talk about her OpenClaw setup, shared what it has been like moving from Claude models to ChatGPT.
Oh and if you haven’t watched Claire’s interview with Lenny, I highly recommend it, probably the best interview I’ve seen from a real OpenClaw power user, and a ton of interesting stuff shared in it.
Okay, but now let’s talk about what happens next, for you, me, and many other OpenClaw users.
First I’ll share what changes for me.
Nothing.
I use Claude with OpenClaw through the API. I always have. But, I don’t use Claude for everything. Personally, and I think many people would agree with me here, Claude is overkill for a ton of stuff that you’re likely doing, or going to do with OpenClaw.
What I think I’ve learned over the last 24-hours is that most people have no idea that Multi-Agent Routing exists in OpenClaw.
This is in the official docs, I can’t recommend it enough.
When I told people on X last night, that I use OpenClaw with Claude, through the API, I got the same comment over and over again. You must be spending a fortune, like $7,000/mo - you’re crazy!?!
I use Claude models for only a fraction of what I do with OpenClaw. Turning on lights in my house, yeah - don’t need Claude for that. I use a combination of models, based on the agent, and what it’s supposed to do. And yes, I use local models for tasks that local models can handle, and that’s at a cost of $0.
As for giving your OpenClaw agent more personality, Peter has a great tweet about this, from back in February, that I highly recommend.
Okay, but I know a lot of people don’t want to switch from OAuth to the API, and suddenly have to manage multiple model orchestration with OpenClaw, I get it. One thing I will say is, from a security standpoint, be careful.
This is your official warning, frontier models are the only models I know of, that can actually truly protect you from prompt injection attacks. If you don’t know what prompt injection is read about it, here’s a solid, free resource:
Unfortunately, I think a lot of people are going to move from Claude models, to another model, that is cheaper, or free, but dramatically increases the risk of getting hacked. Please, don’t rush into anything, take the time to learn before making moves, especially if you’ve given OpenClaw access to things like your email.
Now, for those of you just ready to make the move, and you want to stick with frontier models, GPT 5.4 is probably the direction you’re going to move in. To do this, you’ll likely want to do a little agent tuning, and I think this article does a great job of walking you through the process:
Last but not least, I wanted to end with a hard truth, but an important one. If you want to get value out of OpenClaw, spending just a few hours with it here and there isn’t going to work. OpenClaw is incredibly powerful, but it requires a lot of time, blood, sweat, and yes - tears, to really leverage it properly.
I see way too many people setup OpenClaw in an hour or two and then say, “this is stupid, nothing works right,” and then they abandon ship. There’s a reason why Jensen, the founder and CEO of NVIDIA and people like Mark Andreessen see OpenClaw as a major technological shift. Mark literally says it is “one of the 10 most important software breakthroughs.”
So not learning how to use OpenClaw, or having only used it a few months ago for a few hours, yeah - that just means you’re going to miss the boat on something big. But you have to put in the time, otherwise your expectations just aren’t set correctly.
Ryan says it well:
Okay, that’s all I got on a Saturday morning. Claw on 🦞
