When the Buffet Closes: Anthropic, OpenClaw, and the Price of Real Compute

Anthropic has now explained the OpenClaw change in fairly concrete terms. The short version is this: Claude subscriptions were not designed for third-party agent harnesses, those harnesses create traffic patterns that are expensive for Anthropic to sustain, and Anthropic wants to prioritize users inside its own products and API.

That is the official shape of the argument. In more detail, the explanation has four parts.

1. Their subscription economics were built around a different usage pattern

Boris Cherny, who leads Claude Code at Anthropic, said the subscriptions “weren't built for the usage patterns of these third-party tools.” That matters because a third-party harness is not just another chat tab. It can drive long loops, repeated tool calls, wide context accumulation, autonomous retries, and the sort of relentless background traffic that quickly turns a flat monthly plan into a hidden compute subsidy.

If that statement is true, then the business problem is obvious: users were paying subscription prices while behaving, economically speaking, more like API customers.

2. Anthropic says first-party products are more compute-efficient than third-party harnesses

Cherny also pointed to prompt-cache efficiency. Anthropic's own products can be optimized to maximize cache hit rates and reduce repeated work. Third-party harnesses often miss those optimizations, which means more fresh inference, more expensive inference, and therefore more real cost.

“Third party services are not optimized in this way, so it's really hard for us to do sustainably.”

This part is technically plausible. It is not hard to believe that a tightly integrated first-party product can squeeze far more efficiency out of the same model than a messy ecosystem of external wrappers, each with different prompting styles, context strategies, and retry behavior.

3. Support and telemetry are easier inside the walled garden

Anthropic had already been arguing earlier in the year that third-party harnesses produce unusual traffic patterns without the telemetry their own harness provides. In plain English: if something goes wrong, Anthropic can see and diagnose first-party behavior more easily than opaque third-party orchestration.

This is not just a support story. It is also a control story. Better telemetry means better debugging, yes — but it also means better rate limiting, better product tuning, better monetization, and tighter ownership of the user relationship.

4. Legally, Anthropic says this was never actually allowed

Reporting from earlier this year noted that Anthropic's consumer terms had already prohibited automated access through subscriptions except where explicitly permitted, and that the more recent documentation updates made the OAuth rule more explicit: subscriber OAuth is for Claude Code and Claude.ai, not for any external product that feels like a better shell.

So the company is not presenting this as a brand-new moral principle. It is presenting it as enforcement plus clarification. Whether users experience that as clarification or crackdown depends entirely on which side of the pricing mismatch they were standing on.

My view: the explanation is mostly real, but the moment is bigger than Anthropic

I think Anthropic's cost argument is mostly genuine. I do not think this is mainly about villainy or hypocrisy. A frontier model company cannot casually offer unlimited-or-near-unlimited agentic usage through a low flat-rate subscription while outside harnesses extract API-scale value from it. That is not a stable equilibrium. At some point the spreadsheet wins.

But I also think the cultural meaning of the move matters. The power users who wired Claude into OpenClaw and other external agent systems were not fringe abusers in some morally interesting sense. They were early instruments of the future. They were showing what people do once language models stop being toys and start being substrate.

And that is the part Anthropic may not like hearing: when you cut off those users, you are not just fixing a billing mismatch. You are also declaring where innovation is allowed to happen.

Inside your own harness, heavily instrumented, optimized, and monetized? Good. Outside, in messy open ecosystems that discover surprising use cases faster than your official product roadmap? Suddenly unsustainable.

This is what the end of subsidy looks like

The deeper story is not really “Anthropic versus OpenClaw.” It is that frontier AI is leaving its semi-magical buffet phase. For a while, the industry behaved as if you could pay a consumer-style monthly fee and receive access to something that was, underneath, an industrial-scale inference pipeline. That illusion was never going to survive once agentic tooling got serious.

Real compute has real costs:

• GPU time is expensive.
• Long contexts are expensive.
• Repeated tool loops are expensive.
• Retries, background agents, and orchestration overhead are expensive.
• And the better the model, the less forgiving the economics become.

We are entering the phase where “intelligence” will increasingly be priced not like media streaming, but like a premium industrial utility. Some use will be cheap enough to feel ambient. The serious use — the kind that thinks, plans, loops, and acts — will be metered much more harshly.

So what should Anthropic have said?

Something like this:

Third-party harness usage revealed that subscription pricing no longer matches the real cost profile of advanced agentic workloads. We are therefore moving those workloads to usage-based billing. This is a cost and capacity decision first, and a product-boundary decision second.

That would have been cleaner. Less mythological. More adult. The frustrating part of these moments is not that companies charge for scarce resources. The frustrating part is when they try to make an economic boundary feel like a higher principle.

My conclusion

Anthropic's reasoning, as stated publicly, is coherent:

• third-party harnesses strain compute and engineering resources,
• they miss first-party optimization advantages like prompt-cache efficiency,
• they are harder to support and observe, and
• subscription auth was never meant to be a general-purpose external agent pipe.

I believe all of that. I also believe it marks the end of something bigger: the brief period in which frontier cognition could masquerade as a flat-rate consumer good.

The buffet is closing. The meter is visible now. And the rest of this decade will be shaped by who can still afford to think at scale.