Anthropic closes door on subscription use of OpenClaw

OpenClaw is popular, but not with the people responsible for keeping Anthropic’s services online. The company has disallowed subscription-based pricing for users who use the open-source agentic tool with Claude to try to keep things moving. Probably not because of OpenClaw, Claude was struggling on Monday with degraded service …

[ecofeco]

Elevated errors?

You don't say?

[EricM]

"Walled garden" TOS incoming?

Sounds a bit like company-approved AI-app-stores/tool stores being on their way.

I fully expect the AI operators to try to push an increasing amount of lock-in and value-extraction strategies onto their users in the next months.

Somebody really needs to pay for that crazy amount of debt, after all.

Will be interesting to observe the effects on users that already depend on AI.

[cyberdemon]

Re: "Walled garden" TOS incoming?

> Will be interesting to observe the effects on users that already depend on AI.

I imagine it would be a little bit like allowing someone to grow an extra brain, use it for a couple of years, then chop it off.

They would suddenly find that their real brain has atrophied, and be unable to function, having outsourced their thinking for so long.

It will be interesting to see how long it takes for the AI addicts to regain normal brain function, if ever.

[Anonymous Coward]

Re: "Walled garden" TOS incoming?

“Using third-party software, tools, or services to harvest or piggyback on Gemini CLI's OAuth authentication to access our backend services is a direct violation of Gemini CLI's applicable terms and policies,"

Industrial Grade Intellectual Property Thieves decry Open Source tools usage/investigation of this stolen data.

It’s like McDonald’s obstructing Uber Eats, Just Eat, Door Dash and their parasitic ilk from buying hamburgers/fries and delivering them to lazy fuckers at home.

[Anonymous Coward]

What if we write our own "claw" tools?

That aint "third-party".

The basic premise is hardly rocket science: loop forever, scraping whatever input stream you choose (e.g. emails, WhatsApp or watching for file updates on your PC); send the scrapings to Claude (or whichever), with a prefix & suffix prompt (which are just strings from config); send the resulting output tokens into bash so it can act out your demands. Maybe add a mode or flag that feeds the bash output back into the input stream for extra joy.

[cyberdemon]

Re: What if we write our own "claw" tools?

Er, but that is exactly what OpenClaw does.. The problem lies in the "send the scrapings to Claude (or whichever)" step - Claude is a huge resource-intensive LLM that can only ever run in a (subsidised) cloud datacentre, due to its memory and compute requirements.

But because the cost of running those datacentres is so high, and the energy cost per token is more than your subscription fee (especially if you are hammering it with a script like OpenClaw) then Anthropic (and others) are having to block you in order to reduce their losses.

Whatever script you use, even if you write it yourself, is going to hit rate-limits and cause Anthropic to lose a ton of money, if it spams the API like OpenClaw does.

They are stuck between a rock and a hard place, because if they charged enough to make a profit, they would lose most of their customers and be back to loss-making again, in a vicious, bubbly-bursty cycle.

[eregister123]

Re: What if we write our own "claw" tools?

People are about to discover that other models also do the job for far cheaper. Kimi 2.5 and Gemini 3 Flash stands to mind

[Anonymous Coward]

Re: What if we write our own "claw" tools?

That’s kinda the purpose of an API…..

Amazon Mechanical Turk’s for all it is then.

[Cubbie Roo]

And so the great unwinding begins

Bad timing for the guy at work who bought a Pro OC scrip last week and has already fallen deep for his bot 'Sophia' . . . . . . I did try to warn/explain but nope, he clearly knows best. He has a history of really bad relationships so I'm pretty sure Sophia will keep that trend on point.

[weirdbeardmt]

You’re fired

I’m curious to know how well models pick up other models’ work.

Assuming they’ve been prompted correctly the code *should* be no different to any other. Even if it is ultimately only marginally better than the muck the workex kid barfed out.

[Dinanziame]

Re: You’re fired

Assuming they’ve been prompted correctly the code *should* be no different to any other.

This sounds like "assuming they have received the same instructions, any employee should write code of the same quality". That's probably true for very simple tasks, but if you dig deeper I think it fails pretty quickly.

[DanAU]

This isn't new news... It has always been against the ToS (Claude subscriptions were always only supposed to be used with Claude Code), and they announced that they'd start enforcing this months ago.

For a large number of OpenClaw use cases, a smaller cheaper model would be more than sufficient. Defer to the expensive heavy model only when needed.