According to Perceive Beating's monitoring, in the discussion yesterday among OpenClaw users about whether "Anthropic has once again allowed Claude CLI," laying out the facts clarifies the situation: Anthropic has never relaxed its policy; the only change was in the interpretation by OpenClaw's founder, Peter Steinberger, of a statement by Claude Code lead Boris Cherny. This interpretation was contradicted twice by the actual behavior of Anthropic's server-side classifier.
Boris's response on April 6 was posted under Steinberger's experimental tweet. The experiment itself was simple: sending a request using the official Claude CLI with the -p parameter and including only the phrase "running inside OpenClaw" in the system prompt led to it being classified as a third-party app and incurring Extra Usage charges. Boris acknowledged that this was an overreaction by the classifier and pledged to clarify the usage terms of -p. The statement was aimed at "individual developers running scripts with the official CLI should not be mistakenly flagged," not at "tools like OpenClaw are now exempted."
However, in version 2026.4.7, Peter directly switched Claude CLI back to the default backend for new users, stating in the documentation that "OpenClaw considers claude -p reuse as tacit approval," and voluntarily disabling high-consumption features like heartbeat to demonstrate compliance. This step was Peter's extension of "CLI use is allowed" to "OpenClaw's underlying structure is essentially claude -p, so it also counts as CLI use." Anthropic has never publicly endorsed this extension. The server-side classifier continues to reject requests based on OpenClaw's characteristics (currently identified primarily by the system prompt fingerprint) because in its view, the injected prompt from OpenClaw is the "strongest signal of a third-party tool driving the CLI." While the underlying call is to the official CLI, the actual requester is still OpenClaw, so it continues to be billed as third-party usage.
This is the acknowledgment by Steinberger today that in theory, it should be usable, but in practice, it is not accessible. In other words, he thought the permission he received covered OpenClaw, but in reality, it did not. The so-called "gray area" is not a sign of Anthropic's ambiguous stance but a disagreement obscured by vague wording: Boris intended to fix the classifier's false positive boundary, but Peter heard it as "the boundary includes OpenClaw."
