"Claude CLI lifts OpenClaw ban" misunderstanding was actually a unilateral misinterpretation by Peter

According to Beating Monitoring, yesterday OpenClaw users discussed whether “Anthropic has once again allowed Claude CLI,” and clarifying the facts makes it clear: Anthropic has never relaxed its policy; the change is only in how OpenClaw founder Peter Steinberger interprets a statement from Boris Cherny, the head of Claude Code, which the classifier on Anthropic’s server has denied twice through actual behavior.

Boris’s reply on April 6 was posted under Steinberger’s experiment tweet. The experiment itself was simple: sending a request with the official Claude CLI’s -p parameter, and only including the phrase “running inside OpenClaw” in the system prompt, was classified as a third-party application and charged from Extra Usage. Boris admitted this was an overreaction caused by classifier abuse and promised to improve the usage terms for -p. This statement was aimed at “individual developers running scripts with the official CLI should not be mistakenly flagged,” not “tools like OpenClaw are exempt from this.”

However, on April 7, v2026.4.7, Peter directly restored Claude CLI as the default backend for new users, stating in the documentation that “OpenClaw’s reuse of claude -p is considered implicitly approved,” and proactively turned off high-consumption features like heartbeat to demonstrate compliance. This step extended his view of “CLI use is allowed” to mean “OpenClaw’s underlying system is essentially using claude -p, so it also counts as CLI use.” Anthropic has never publicly endorsed this extension. The server-side classifier continues to reject based on OpenClaw features (currently mainly the system prompt fingerprint), because the prompt injected by OpenClaw is seen as the strongest signal that “third-party tools are driving the CLI.” Although the request is made via the official CLI, the actual initiator remains OpenClaw, so it is still billed as third-party usage.

This is what Steinberger publicly acknowledged today: “In theory, it should be usable, but in practice, it isn’t.” In other words, he thought the permission granted covered OpenClaw, but it turned out it did not. The so-called “gray area” is not due to Anthropic’s ambiguous attitude but a misrepresented divergence: Boris wanted to fix the classifier’s false positives, while Peter interpreted it as “OpenClaw being within the boundary.”