So you’re paying for a Claude subscription, using it the way you want, and then boom. Less than 24 hours notice and suddenly your setup is against the rules.
That’s exactly what happened last Friday at 4 PM. The creator of this video, Matthew Berman, broke down the whole situation while literally on vacation because it was that wild.
🔥 Here’s what went down:
Anthropic sent an email saying that as of April 4th, using third-party harnesses like OpenClaw with your Claude subscription is now officially against their terms of service. They gave users roughly 19 hours to make changes. That’s it. And they specifically named OpenClaw in the email.
To soften things up, they offered full refunds if you want to cancel. Boris Cherny, head of Claude Code, confirmed that subscriptions just weren’t built for the usage patterns these third-party tools create.
📊 Why is Anthropic doing this?
The answer is simple: they’re in a massive GPU crunch. Here’s what the expert laid out:
- Anthropic’s revenue run rate jumped from $9 billion at end of 2025 to $30 billion now, finally surpassing OpenAI
- Their status page is covered in red. Total uptime for claude.ai sits at 98.77%, which sounds decent but anything below 99% is rough for a production service
- They tried carrots first (2x usage during off-peak hours) and sticks (faster session limit burns during peak 5-11 AM Pacific). Neither was enough
- A Cursor report suggests a $200 Claude subscription actually costs Anthropic around $2,000 in credits. That math doesn’t work at scale
🔄 The switching cost is basically zero
Here’s the part I found most interesting. The creator showed how he swapped all his Claude models in OpenClaw to GPT 5.4 in about 3 minutes. Just told OpenClaw to switch everything over, including the optimized prompt files. Done.
One pro tip from him worth highlighting: keep multiple variations of every prompt file optimized for different models. A prompt tuned for Opus 4.6 looks completely different than one doing the same job on GPT 5.4. That prep work makes model-switching painless.
Jack Dorsey even agreed with his take that there’s literally zero switching cost between frontier models right now.
😵 The policy clarity problem
What makes this even more frustrating is that weeks later, Anthropic still hasn’t clarified key questions. Is the Agents SDK allowed with OpenClaw? Is it allowed with extra usage? Nobody knows. Boris Cherny said “no changes to Agents SDK at this time” but that they’re “working on improving clarity.” Meanwhile, there are reports of an overactive abuse classifier blocking even first-party harness use and flagging certain system prompts.
🧠 The real takeaway: go multi-model
The expert’s advice here is spot-on. If you’re building anything with agents, you need a multi-model strategy. Not just multiple frontier models, but local models too. Classification, data extraction, summarization can all run on open-source options like Gemma 4 or Qwen 3.5. Save the frontier models for the heavy lifting like orchestration, planning, and coding.
Meanwhile, OpenAI is rolling out the welcome mat. They brought in Peter Steinberger (OpenClaw’s creator), keep resetting quotas generously, and the latest OpenClaw version apparently makes GPT 5.4 feel natural to use.
Want the full breakdown with all the receipts and screenshots? Check out the full video for every detail.