Update, March 19: This article has been updated to reflect that Anthropic has not made a formal announcement regarding OAuth deprecation. Based on code commits, deprecation appears underway, but the rollout is inconsistent and workarounds remain functional for some users.
Anthropic appears to be deprecating OAuth token access — the method that allowed OpenClaw and other third-party tools to authenticate using Claude Free, Pro, or Max subscription credentials.
The change was identified by Dave Swift on March 19 through upstream code commits. Anthropic has not made a formal announcement, and the deprecation has reportedly affected some users while others retain access. ShareUHack published setup guides outlining the three authentication methods that still work.
The Cost Gap
OAuth access was the primary cost arbitrage for budget-conscious OpenClaw operators. A Claude Max subscription costs $200/month and provided effectively unlimited usage when accessed via OAuth tokens. The equivalent workload through Anthropic’s API — where Claude 3.5 Sonnet charges $3 per million input tokens and $15 per million output tokens — routinely exceeds $1,000/month for agents running continuous workflows.
For operators running OpenClaw as an always-on personal assistant or autonomous worker, this represents a significant cost increase overnight.
A Temporary Fix Exists — For Now
Swift’s post documents a configuration workaround that temporarily restores OAuth access. The specifics involve modifying the OpenClaw configuration to route authentication through an alternative token path. But Anthropic’s commit signals full deprecation is incoming, making any workaround a stopgap rather than a solution.
What Operators Are Weighing
The deprecation resharpens a question that’s been simmering in the OpenClaw community: is Claude the right default model for cost-sensitive deployments?
The alternatives have gotten materially better since OpenClaw’s early days:
- Cheaper API models like Kimi K2.5 offer comparable agentic capabilities at a fraction of Claude’s per-token cost
- Local models via Ollama (Nemotron Nano 4B, Qwen) eliminate API costs entirely at the expense of capability
- OpenAI’s GPT-4o sits between Claude and the budget options on both price and performance
For operators whose agents perform routine tasks — calendar management, email triage, file organization — the quality gap between Claude Sonnet and a cheaper model may not justify the cost premium once OAuth is gone.
For operators running complex multi-step agents that depend on Claude’s instruction-following and tool-use capabilities, the API switch is a cost of doing business — but one that changes the economics of running OpenClaw as an always-on system.
The Bigger Picture
Anthropic’s move follows a pattern across frontier model providers: closing off indirect access paths that allow subscription pricing to subsidize API-grade usage.
The practical effect is that the era of running frontier-model-powered agents for $200/month is ending. The agents still work. They just cost what they actually cost.