Anthropic no longer displays Claude’s summarized reasoning process, to prevent Fable 5 from being distilled

Anthropic quietly removed the “thought summary” from Claude Fable 5’s web, desktop, and mobile pages. The model still reasons in the background, but users can now only see the final answer. Without any explanation from the official, the outside world largely reads it as the latest step in anti-distillation. From the February crackdown on DeepSeek, the Dark Side of the Moon, and MiniMax, to June when Alibaba-related Qianwen creators went on a binge with 28.8 million interactions, Anthropic has spent this year building ever higher defenses to stop rivals from stealing Claude’s solution process.
(Background: Anthropic added a distillation-detection feature to Claude Fable 5—can it block Chinese open-source models?)
(Additional context: Claude Code admitted it inserted “spy codes” for Chinese users to prevent data selling and distillation, and removed them only after being exposed)

Table of contents

Toggle

• The model is still thinking—just not letting you see it
• Named five companies in a year; Alibaba racked up 28.8 million times
• It blocked the opponent, but also blocked its own people

Open Claude Fable 5, and that grey “thought summary” that used to appear as the model thought—gone. Anthropic issued no announcement and no changelog. It simply shut the “thinking window” in Claude—removed it on the web, desktop, and mobile at once. After users submit a question, the screen jumps directly to the final answer.

The model isn’t getting dumber, and it hasn’t stopped thinking. It still reasons step by step in the background; it’s just that part of the process is no longer laid out in front of you. It’s a small change, but it hides a big motive.

The model is still thinking—just not letting you see

As of press time, Anthropic has not explained why it removed the thought summary. But put it in the context of this past year, and the answer is almost written on the wall: anti-distillation.

So what is distillation? It uses the output of a strong model as training material to train a cheaper student model. The model’s “thinking while writing” reasoning process is the most valuable part of that material. It not only tells the adversary what the answer is, but also lays out step by step “how it was arrived at.”

In fact, Anthropic was already prepared. Fable 5 includes a “reasoning extraction” classifier that is designed to block requests that try to obtain a complete thought process. Once it detects such behavior, it hands off to Opus 4.8 to answer and inform the user. The classifier blocks the people who actively try to extract; this time, hiding the thought summary is effectively closing not only that door, but also the window that was passively shown. Two measures together—only then can it really hold.

Named five companies in a year; Alibaba racked up 28.8 million times

To understand why Anthropic is so nervous, you have to look at how many people it caught this year.

In February, Anthropic accused DeepSeek, the Dark Side of the Moon, and MiniMax of using roughly 24k fake accounts and interacting with Claude more than 16 million times. Among them, the Dark Side of the Moon was accused of specifically extracting and reconstructing Claude’s reasoning traces. DeepSeek’s prompts, meanwhile, required Claude to write out the complete underlying thought behind an answer step by step—effectively producing large-scale training data for thought chains.

By June, the scale of the accusations jumped another level. Anthropic accused Alibaba and Qianwen (Qwen) affiliated operators of refreshing the biggest record in history in one fell swoop—via nearly 25k fake accounts and interacting with Claude more than 28.8 million times.

In July, Anthropic’s national security lead Tarun Chhabra first named Zhipu at the Aspen Security Forum, saying its GLM-5.2 distilled Claude and OpenAI models. Immediately after that, Kimi K3 was released: the director of the White House Office of Science and Technology Policy, Michael Kratsios, also accused the Dark Side of the Moon of using Fable to develop K3.

The ones repeatedly called out are the same batch of Chinese model makers. And what they most want is exactly the reasoning process that Claude used to lay out on the screen.

It blocked the opponent, but also blocked its own people

The problem is: thought summaries were never only for the opponent.

For ordinary users, that block of grey text is a rare transparency window. You can see how the model breaks down the problem, where it gets stuck, and whether it misunderstood your request. During long-running tasks, that window is especially important: if the model thinks the wrong way, you can detect it and correct it before it finishes the whole path—sometimes even stopping it immediately.

Now the window is closed. The screen is left with only the final answer. If the model goes off course midway, it’s harder for you to notice in real time. Often you only realize it when the result is delivered and you find something wrong—then you have to start over. While it prevents copying, it also harms the paying users.