Claude ended subscriptions in OpenClaw. Try these alternatives.

After Anthropic kicked OpenClaw users off its subscription service, @Shpigford got methodical about finding an alternative. These screenshots show which alternatives are worth trying and which suck. (It all comes from our recent YouTube interview.) Claude's "get off my lawn" email On Friday, Josh got an email from Anthropic telling him to stop using his Claude subscription in OpenClaw. They're now insisting that users switch to (more expensive) API access. The price difference would be huge: API: $1,500/month (minimum) Subscription: $200/month He decided to look for an alternative. How he ran his test Josh uses Discord to communicate with his OpenClaw agent. So he created a new chat in Discord for each one of the models, to compare them to Open 4.6 (his previous daily driver). Then he asked each one the same thing. To connect with each one of the models, he used Open Router. The only exception was GLM 5.1, which needed a direct connection from Z.ai. The question he asked: This is a typical question he has for OpenClaw, one that involves understanding both context and tools. Opus 4.6: 🟢 the clear winner This is exactly what he was looking for. It knew that the playbook was in Notion, without being told. It gave him the next item, just as he asked for. But look at all the unasked-for relevant context it added. Opus 4.6: Knew that the document he referenced was in Notion Told him that the doc was based on a March 23 deep strategy session Gave him other tasks from that playbook suggested which tasks were highest-leverage, lowest-effort Explained its reasoning Offered to help with the next step Sonnet 4.6: 🔴 didn't find it It didn't know that the doc was in Notion. It asked for "a quick refresher." Gemma 4: 🔴 clueless It said, "I'm not seeing any prior work" on this playbook. Kimi K2.5: 🔴 couldn't find it It showed how it was thinking, but couldn't find the document. GPT 5.4 mini: 🟡 just the answer It was a "just the facts" response. It's clear that it knew to go into Notion and to find the right playbook. It got him the next task, just as he asked for it. But it didn't put it in context. It didn't offer to help. It didn't advise him. This is the frustration people on X have had with Open AI models. GLM 5: 🔴 didn't get it It showed its work -- in detail -- but in the end it couldn't figure out that the document was in Notion. It didn't even know to search for it there, even though Notion is clearly documented as one of his OpenClaw's tools. Qwen 3.6:🟡 needs handholding It found the Notion integration. It even discovered from session notes that the data was in Notion. But it didn't know that it had Notion access. It didn't try to log in. It needed more handholding. So yes, Josh could say "go check Notion," but it's not the kind of guidance he's used to giving his OpenClaw. GLM 5.1: 🔴 heartbreaking This one broke Josh's heart. You can see it in the video. It worked yesterday. He saved it for last so he could show how well it worked. But it faceplanted. Couldn't find the doc. Bottom line For @Shpigford, Opus 4.6 is expensive (at least $50 per day, for his use), but it consistently gets the right answers, doesn't need handholding, and just "gets it." But in our video, he showed how you can easily test it for your use cases. The full video is the latest interview on my YouTube channel, here. Next Follow @AndrewWarner for more AI builder interviews like this. Follow @Shpigford to see what he ends up doing with his OpenClaw.