Why people dislike Anthropic

Just six months ago, we all loved Claude Code and recommended it to each other. We recommended it so much that, by inertia, everyone still continues to think it is something good. Let's reconstruct the history, because over this time there have been several turning points that gradually made people turn away from them. Today is another one.

In short. There are Claude Code, Claude Web, and Claude Desktop — Anthropic's first-party products. They also have the Agent SDK, which allows other applications to use the same tools, loop, and other Claude Code features. Claude Code has both a TUI and a non-interactive mode, claude -p "make it beautiful", which allows other applications to go ask Claude Code something without a user interface. This mode is built on top of the Agent SDK.

And about money. In all these tools you can pay for tokens used if you work through the API. Or you can pay for a $20 / $100 / $200 subscription and use it within certain limits. If you exhaust the limits, there is an option to pay for tokens beyond the subscription. Subscriptions are, of course, much cheaper than tokens, because they are used more for marketing than for revenue. Anthropic's revenue is a consequence of this image of being the best agent provider, which enterprise has bought into, paying thousands of dollars for tokens.

If you combine the first and the second, it turns out that for the entire lifetime of Claude Code, regular mortals could use a cheap subscription together with third-party tools: OpenCode, OpenClaw, Hermes, and even their own Ralph Loop in bash — all of them use this. It is very convenient, but Anthropic really does not like it. And they obviously regret that they gave people this option in the first place.

The whole point is in the history of Anthropic's actions. Let's even skip, in this text, the fact that subscription limits keep getting smaller and smaller. That can be justified as an artificial limitation while there is not enough compute. The main thing is that they quietly suppress any even slightly hyped product built on top of this. First OpenCode fell, because they simply forbade it from authenticating the way it did. Then they came after ClawdBot with lawyers, hiding behind the similarity of the name, though of course what they really disliked was subscription abuse by automated agents. Then they explicitly wrote in the terms of service that a subscription cannot be used anywhere except first-party applications. Agent SDK and claude -p have always been first-party, so on Twitter they had to justify themselves by saying that the TOS is nonsense and that you can still use them as before.

One of the funniest and saddest parts of this story happened last month. Users noticed that even inside Anthropic's own harness, something strange was happening. Sometimes Claude Code responded with errors to messages mentioning the tools named above. While fighting this, they added even more junk both to the system prompt and to their code, which hurt quality. And then the cherry on top: they quietly started billing for it as extra tokens. Inside Claude Code. If they saw the word Hermes somewhere in your commit history, for example, you were charged money. Even if you had no idea that such software existed. This is about as wrong as it gets.

And now they are separating billing inside their applications and outside them. Starting June 15, for a month they will give you credits equal to the cost of your subscription, and any call through Agent SDK or claude -p will spend these credits like API usage, without affecting your limits. At first glance, it sounds as if they are giving you another $100 on top of your $100. But in practice it more likely means that even for your own script, or your favorite orchestrator with Claude under the hood, the limit has become about twenty times smaller.

In short, you really have to try hard to communicate with end users this badly and act this inconsistently. That is why people dislike them.