What's Anthropic up to?

People are upset because Anthropic disconnected itself from OpenClaw. What’s going on?

People are upset because Anthropic disconnected itself from OpenClaw. What’s going on? Is it some kind of petty sabotage? Or are there good reasons for Anthropic’s behavior?

When OpenClaw first took off, users were allowed to hook it up to their Claude Max subscription and use the best models in the world for a discounted price. But that price was based on several assumptions, chief of which was that you weren’t going to stay awake 24 hours per day prompting the LLM. OpenClaw spiked Claude usage to extremes, and as a result, Anthropic has forced OpenClaw users to go through the API instead, where the cost is much higher.

There’s nothing evil or petty about what Anthropic been doing: they are managing the load on their servers and they’re letting the market set the price for undifferentiated tokens. They have the best model, and people are willing to pay a premium to get that model’s performance. By setting the price high, they ensure that they can deliver reliable good performance for the people willing to pay for it.

The decision to lock subscribers into Anthropic’s native software is also strategic. They have realized that the jagged frontier means that models will eventually become commodities, with a few specialized exceptions. The value will instead accrue to the layer of intelligence that drives the models. This layer is called a harness. Claude Code is a harness, as is OpenClaw, and OpenAI’s Codex, and so on. This is where the real competition is. There is no reason to subsidize the use of their best models on behalf of a competitor.

Claude’s native tools are also likely to be more efficient token users than external harnesses. When Anthropic sells tokens at full price via their API, they don’t care if you are efficient. But when they subsidize in on a subscription, they care a great deal. Most of my OpenClaw API calls are 20,000 tokens or more, because the harness is passing along a lot of context for the model to use, every time. And I have to make my own choices about which model to use for which task.

I assume Claude Code & Cowork are taking advantage of system prompts rather than repeating common bits of context (why pass the same preamble for all Claude code users? Why not just tune a model specifically for their own harness to make some universal context unnecessary?)

I also expect they are using intelligent routing, breaking requests into smaller chunks that get assigned to Sonnet or Haiku or even customized models that are unavailable to the public via the API, based on a sophisticated understanding of what the user is asking for. All of this would be managed by the harness, and therefore it is probably cheaper to Anthropic when you use their software vs something like OpenCode.

So Anthropic is making good strategic choices to tempt people to use their proprietary software, which ends up being a benefit to both the user and to the company because the vertical integration allows them to offer a better experience at a lower cost.

On a recent More or Less podcast episode, the CEO of Vercel proposed that subscription based pricing for AI is on the way out and that it won’t work long term. The argument is that AI has to have usage based pricing like electricity because subscriptions require some users to pay for unused subscription capacity, and AI is so useful that everyone will max out all their tokens.

But I think most people are not going to behave like the enthusiasts, and the certainty of subscription pricing at $20/mo is valuable enough to the average consumer that the pricing model is here to stay. In fact, Anthropic is showing us the future already: a base layer subscription with the option to pay for overages.

Lastly, I think how Anthropic’s chose to release their latest model (called Mythos) demonstrates some really interesting foresight. I recently watched The AI Doc: Or How I Became an Apocaloptimist, and I think the filmmaker does a great job of fairly representing wildly different camps and how they view AI. The documentary shows that people who worry are worried about the incentives in the AI race: that open competition leads to recklessness in releasing potentially dangerous intelligence.

But what Anthropic has done is release Mythos to a limited set of customers. They did this because they say that Mythos is so powerful that it would be dangerous to make a broad release at this time. That means that anyone else who makes a broad release of their model is admitting that their model is not as good as Mythos.

So now we have an incentive where the leading providers of the most powerful models are motivated to control the release of those models in order to signal how intelligent they are. I find that very encouraging.