Anthropic Subscription Credentials and OpenClaw: A 2026 Policy Dispute

Matthew Berman reacts to Anthropic documentation saying OAuth tokens obtained from Claude Free, Pro, or Max accounts were intended for Claude's own supported products and not other tools. He believes the restriction affected OpenClaw users, although the wording he shows does not name OpenClaw. This is a restriction on a credential and subscription use, not a blanket ban on using Claude models through the API.

Berman describes OpenClaw as a locally runnable personal assistant able to use different model providers. He recounts its earlier names, Clawdbot and Moltbot, and says trademark concerns led to a rename. The policy discussion takes place in February 2026; terms and supported integrations can change.

He explains why the distinction matters to him: a subscription provides usage within plan limits, whereas API requests are separately metered. Berman says he moved from a $20 subscription to higher Max tiers as his OpenClaw use grew. His personal spending and effective discount do not establish a universal price comparison for every workload.

Using the prices he cites for Claude Opus 4.6, Berman calculates that 50,000 uncached input tokens at $5 per million would cost $0.25 before output charges. His configuration sends a large context even for a simple greeting. Actual charges depend on input and output volumes, caching, context tier, provider pricing, and the configuration; the example is not a minimum cost for all users.

The displayed documentation also mentioned the Agent SDK, prompting confusion. Berman reads an Anthropic employee's clarification that existing use of the Agent SDK and Max subscriptions had not changed, followed by a distinction between local experimentation and building a business with API credentials. The video itself acknowledges ambiguity, so its reading should not substitute for checking the applicable current documentation and terms.

Berman argues that many OpenClaw users are individual tinkerers who already pay for subscriptions, and he dislikes the restriction. He speculates about Anthropic's economic incentives and compares the reaction with statements he attributes to OpenAI about its own subscription access. These are his account and interpretation of a rapidly changing policy situation, not definitive guidance on present permissions.

Previously Berman used Claude Opus for much of his setup. He says he moved coding and several analysis tasks to GPT-5.3 Codex, assigned simpler classifications to a smaller model, and used GPT-5.2 for other work, with Claude Sonnet and Opus as backups. The arrangement is his personal snapshot, and he notes that it is more complex than his previous configuration.

He mentions also experimenting with Gemini and closes by inviting discussion. The main tradeoff in his example is between a simple subscription-based workflow and the cost, access conditions, and operational complexity of routing requests among providers.