Users Suspect Anthropic of Secretly Cutting Claude Code's “Effort”

Hacker News is buzzing over a series of posts by a user going by the handle argofowl (@argofowl) on the social network X. He claims that Anthropic is running a covert A/B test that lowers the model's “effort” level in the Claude Code tool.

According to the author, the change happens on the server side rather than in the app itself. Some sessions using the Fable 5 model on Claude Code versions 2.1.236 and newer are allegedly swept into an experiment that “compresses” the effort scale. Older versions and the Opus 5 model, by his observations, are not affected by the experiment. Because this is a selective test, not every user sees the effect.

The author was especially blunt about the setting's behavior after version 2.1.237: as he describes it, the model now treats “high” mode as a value of 10 out of 100 — exactly the number that used to correspond to “low” mode. And yet, he stresses, there isn't a word about this in the changelog.

“I Thought My Code Was Broken”

The emotional core of the posts is about how the stealthiness of the change undermines trust. The author recounts spending an entire day convinced that his own app and working code had broken — and only afterward did he suspect the model was to blame. Hence his simple test: “if ‘high’ feels like ‘low,’ you're in the test group.”

Other participants in the discussion picked up on these complaints. One commenter summed up the main grievance with the practice itself: how are you supposed to work with a company that changes the product right out from under your feet?

What's Worth Keeping in Mind

Everything laid out here is one individual user's observations and grievances, not confirmed data. There are no official comments from Anthropic in the material cited, and the conclusions about specific versions and numerical values were drawn by the author himself based on his personal experience. Even so, the story touches a sore spot for anyone who builds workflows around cloud-based AI tools: any imperceptible server-side tweaks to a model's behavior can create the feeling that it's not the service that's breaking, but your own project. The Hacker News discussion has racked up more than a hundred votes and over a hundred comments.

Source: Hacker News