Anthropic just blocked OpenClaw. Here’s what you need to do immediately
TL;DR
Anthropic killed the old OpenClaw subscription hack — Alex Finn says you can no longer connect the $20 or $200 Claude subscriptions to OpenClaw, so Anthropic now forces API billing instead.
His workaround is a 'brain and muscle' setup — use Claude Opus 4.6 as the orchestrator that plans and verifies work, while cheaper models like ChatGPT 5.4, GLM 5.1, Gemini, Qwen 3.5, or Gemma 4 do the actual task execution.
Alex’s core claim is that Opus is still the only model he trusts to actually finish agent tasks — after testing 'every single model on planet Earth,' he says GPT, Kimi, and Qwen can code or write well but often claim success without actually completing the work.
For average hardware, the recommended stack is Opus API + ChatGPT for coding + a cheap secondary model for everything else — specifically ChatGPT 5.4 for coding, then either ChatGPT Pro, GLM 5.1, or Google/Gemini for research, writing, and web tasks.
For strong local hardware, he shifts the 'muscles' on-prem — keep Opus 4.6 in the cloud for orchestration, run Qwen 3.5 locally for coding on a 512 GB Mac Studio, and use Gemma 4 on a DGX Spark or Mac Studio for nonstop web scraping and background work.
The practical move is to let OpenClaw write its own routing rules — Alex shows a prompt like 'use my ChatGPT subscription for all coding tasks,' which makes OpenClaw update its agent.md file so sub-agents automatically hand coding to ChatGPT while Opus checks the result.
Summary
Anthropic shuts the door on subscription-powered OpenClaw
Alex opens with the bad news: Anthropic has blocked OpenClaw from using Claude’s $20 and $200 subscriptions, which means serious users now have to switch to API billing. He frames the risk plainly — if you run OpenClaw heavily, API usage can explode into hundreds or even thousands per day — then counters with his own example: an eight-hour road trip talking to OpenClaw nonstop that only cost him $10 in tokens.
The big idea: smart brain, cheap muscles
His whole workaround hinges on one principle: if the orchestrator is excellent, the executors don’t need to be. He calls it the 'brain and muscle' system — let Claude Opus 4.6 act as the conductor that decides tasks and validates completion, while cheaper models handle the grunt work. The punchline is that you keep Opus-level reliability without paying Opus prices for every subtask.
Why he refuses to replace Opus
Alex gets emphatic here: Opus is 'the goat' for OpenClaw, and he says nothing else is even at 50% of its performance in this context. His main complaint about rivals like ChatGPT, Kimi, or Qwen isn’t raw intelligence — it’s that they often say they finished a task when they actually did nothing. He makes it a competitive point too: if you’re not using the best model for orchestration, someone else is, and they’ll beat you.
The default setup for most people: cloud-first routing
For people with ordinary hardware — think base-model Mac Mini — his recommendation is straightforward. Keep Opus 4.6 on API for orchestration, use ChatGPT 5.4 via subscription for coding, and route 'everything else' like research, writing, and web search either to ChatGPT Pro, GLM 5.1, or Google/Gemini if you already pay for it. His logic is practical: OpenAI allows this usage, ChatGPT is strong at coding, and GLM gives you a cheap fallback if the $20 ChatGPT tier doesn’t provide enough headroom.
The ideal future state: local models doing nonstop labor
Then he switches to the more ambitious setup for people with serious machines like a 512 GB Mac Studio or DGX Spark. In that world, Opus still stays in charge, but local Qwen 3.5 handles coding and Gemma 4 takes on endless background jobs like web scraping and data collection. What excites him most is the always-on angle — local models can run 24/7 without subscription caps, which he thinks unlocks whole categories of use cases most people still aren’t paying attention to.
How to wire it up inside OpenClaw
The setup section is surprisingly hands-on: run OpenClaw onboarding, paste in your Anthropic API key, and, if you’re committed, buy larger Anthropic token bundles to save money — he says he bought the 1,000 bundle for a 30% discount. Then you can literally ask OpenClaw to connect other subscriptions with prompts like, 'I want to use my ChatGPT subscription for all coding tasks,' after which OpenClaw updates its own agent.md rules so coding gets routed to ChatGPT sub-agents while Opus supervises.
His closing argument: stop comparing this to Netflix
Alex ends by reframing the economics. Yes, this may cost more than the old Claude subscription setup, but he argues you’re effectively paying around $2,400 a year — maybe somewhat more — for an AI employee that works 24/7, which he calls 'the biggest steal of the millennium.' He’s still annoyed at Anthropic, but his loyalty is to performance, and at least at the time of recording, he says Opus 4.6 remains the clear best choice for OpenClaw.
Was This Useful?
Share
Keep Reading
Make Alcreon Yours
Tune your feedFive quick questions, and the feed ranks what matters to you first.Or just get notified
The weekly Echo. Signal worth keeping in your inbox.
Every new piece, announced on X.
Read Next
See all
Playbook
The Retirement Email Isn't a Warning
Model retirements now arrive every few weeks; the config-eval-rehearsal loop turns each deprecation email from a fire drill into an afternoon swap.
Playbook
The Cheapest Model That Passes
OpenRouter lists 400 models behind one API. The fix for choosing isn't a better leaderboard, it's a four-step protocol that ends in a real eval.
Playbook
Cheap Models, Hard Tasks
Most agent workflows route every step to the frontier model by default. The bill scales with how chatty the agent gets, even when most steps don't need that brain.