Anthropic meters OpenClaw use

Anthropic has drawn a hard line around Claude. People who used Claude subscriptions inside OpenClaw and similar third-party agent tools are being told that the flat monthly plan no longer covers that traffic. If they want to keep going, they now have to pay extra on a metered basis or move over to the API. Anthropic’s explanation is blunt: agentic workloads are eating too much engineering capacity and too much compute, and the company wants to protect users inside its own products. That change matters because OpenClaw was not some fringe hack. It became one of the clearest examples of what happens when a frontier model escapes the neat box its maker designed for it. OpenClaw lets people route Claude into a broader automation layer that can run across chat apps, workflows, and local systems. Its own documentation had already warned users that Anthropic subscription access outside Claude’s native tools was risky, and recommended API keys as the safer path. Anthropic has now turned that warning into policy. The economics were always unstable. A subscription works when most users behave like chat users. They ask, they wait, they stop. An agent does not stop. It loops through files, runs tests, retries commands, fans out subtasks, and burns tokens while nobody is looking. Anthropic’s own product pages now read like a catalog of exactly those expensive behaviors. Claude Code is sold as a system that can search codebases, edit across files, run command-line tools, monitor CI failures, and keep going until the work is done. Opus 4.6 is pitched for “complex agentic workflows,” with API pricing starting at $5 per million input tokens and $25 per million output tokens. Even Haiku 4.5, the cheaper model, is framed as fuel for coding and research sub-agents at $1 and $5 per million tokens. Once you see the pricing, the policy change stops looking like a surprise and starts looking like delayed arithmetic. Anthropic had already been tightening access as usage surged. In July 2025, it introduced weekly rate limits for heavy Claude Code subscribers. In March 2026, it briefly doubled off-peak usage outside weekday peak hours, which was less a gift than a visible signal that capacity had become something to schedule. The company had been teaching users, step by step, that frontier AI was no longer a buffet. That lesson is landing just as rivals are trying to turn cheap tokens into a weapon. Chinese firms moved fast after Anthropic’s OpenClaw decision. According to the South China Morning Post, MiniMax and Xiaomi publicly urged users to switch, while Zhipu has been advertising an annual plan priced at $84, far below Anthropic’s roughly $204 Claude Pro subscription. The pitch is simple: if premium American labs are going to meter serious agent use, someone else will promise more room for less money. But even that price war comes with an asterisk. The same SCMP report notes that Chinese companies are arguing in public over whether cheaper access is sustainable at all, because AI agents are driving a global token crunch and compute supply is not keeping up. One Xiaomi-linked researcher praised Anthropic for taking a short-term hit to impose discipline. That is the real story here. The market is splitting into two lanes. In one lane, frontier models are premium infrastructure, metered like electricity. In the other, cheaper substitutes race to absorb the overflow. Anthropic just made that divide impossible to miss, and it did it by cutting off one of the most popular escape hatches.