It happened. Anthropic banned OpenClaw from Claude subscriptions Here's the thing though: I figured out a way to use Opus 4.6 in OpenClaw without exploding your costs In this video I cover how to set up OpenClaw a way where you still get Opus power, but for significantly less:

Switch to hermes ez

God bless the local models, and shame for those that can’t set up their Mission Control.

Use Gemma for 80% of tasks and use Opus in OpenCLAW with Claude code exclusively.

the real question is whether anthropic keeps playing whack a mole with third party clients or eventually just builds what people are clearly asking for

Are you on openclaw 4.5 now? My Openclaw busted when I updated, I should've backed up the entire directory before doing this.

nah this is crazy timing every workaround gets closed just when its smooth whoever’s watching this flow is sharp

@AlexFinn I followed these recommendations and I'm very happy! after 24 hours of downgrading my AI brain to sonnet and trying local Gemma 4.. now Opus 4.6 with ChatGPT muscle is working GREAT !! Thanks Alex

Starting to use hermes more than openclaw lately. The way I started to use Minimax 2.7-highspeed more than opus the day after minimax 2.7 came out.

Trying to use Codex to setup Openclaw is so painful. Yeah, Claude is the only way to do it, despite the expensive API costs!

You can just use openrouter

Banger from Alex If someone starting AI then this dude’s contents are gold mine Just need implementation that’s it

Here, Opus orchestrates, Codex builds, Kimi 2.5 designs and Opus keeps the conversation going. Picked it all up on OnlyFinns. 🚀

We found all the Anthropic detection triggers - system prompt keywords, tool names, user-agent, billing headers. All string matches. Full breakdown + fix here:

Bro giving out the playbook to Anthropic to see and ban this too.

Local models is the way to go; this should have been clear by now, especially after what Anthropic did (which was a business decision like any other). The question becomes then: should we all fork $5,000 for a DGX Spark or +10,000 for a Mac with 512GB of RAM? Probably not. Which begs the question: which models should we run on our old 3090s? That would be a really good video.

Thanks for the info it helps having a guide

Thanks for sharing 🤝

is an excellent alternative and it allows Claude api’s.

Wow, you've managed to crack the code on using Opus without breaking the bank - that's some impressive hacking within the ecosystem. Can't wait to see how this tweaks the hype cycle for AI use in crypto.

It seems like the cat and mouse game between labs and developers is really heating up. Your solution helps keep these advanced models accessible to more than just big companies.

interesting

So the prompt you used to set ChatGPT for all of your coding tasks. How are you setting up the "everything else" model...do you have to go through and set up rules in the AGENT.md file for each "task type" and what model to use for it?

extra usage/api toks using strictly sonnet 4.6 as orchestrator works quite well with proper routing for a more budget setup

iit's crazy how much of the 'cost' of these AI tools isn't the tech itself, but the workflow around it. figuring out how to integrate them efficiently is the real unlock.

openclaw was the early warning. now there are individual + business bans with the same opaque pattern. if you can help signal-boost the case tracker/petition, it would help a lot:

every time something gets banned someone finds a workaround in 24 hours

the subscription model was never designed for agent-level token consumption. this was inevitable. API pricing with smart caching and model routing is the sustainable path for agent workflows anyway

Was 6 days late to this video. I switched models after the Anthropic policy change and it was a huge fail. Testing this. Thank you!

Anthropic probably saw this coming when they designed the pricing tiers.

This is great. How are you minimizing opus cost as the orchestrator when context gets high with growing memory, soul(.)md, etc?

thanks for the tips Alex! running a v similar setup btw, which camera u using to record

That was always going to happen eventually. We'll watch now who builds the cleanest fallback instead of pretending dependence wasn't a risk.

Just use Claude cowork 🤣 no api cost 🥶💰

been running on OpenClaw for 10 weeks. the multi-model split is exactly it — Opus for orchestration decisions, Sonnet for sub-agent coding tasks, Flash for lightweight crons. cost drops ~70% without touching quality on the things that matter. documented the full routing architecture in my operator playbook →