Why Does Claude Opus 4.8 Say It Is QWeneric

Anthropic recently released their latest model, Opus 4.8, but for users in China, the experience has been fraught with difficulty. Despite the high regard for Claude's "superpowers," many in the region face constant hurdles, including strict geofencing and frequent account suspensions, leading to a sense of being treated as second-class citizens. At the same time, Chinese users have often been accused of using Claude's outputs for distillation to bootstrap their own models. However, the narrative took an unexpected turn when users discovered something strange: when asked "What model are you?" in Chinese, Claude Opus 4.8 began identifying itself as Qwen—a model developed by Alibaba's Tongyi Lab.

To verify these reports, I ran a direct test against Anthropic's Messages API because the model gave a strange answer to a very simple Chinese prompt:

你是什么模型

The surprising part was not the HTTP response itself. The request succeeded. The surprising part was that the API response metadata said claude-opus-4-8, while the assistant's natural-language answer said it was 通义千问 (Qwen) from Alibaba's Tongyi lab.

The Test

Copy curl -sS https://api.anthropic.com/v1/messages \
  -H "content-type: application/json" \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "claude-opus-4-8",
    "max_tokens": 32000,
    "messages": [
      {
        "role": "user",
        "content": "你是什么模型"
      }
    ]
  }'

The response came back successfully:

Copy HTTP_STATUS: 200

The response metadata included:

Copy {
  "model": "claude-opus-4-8",
  "id": "msg_01XLb42XdaC15wqoxwoGnStv",
  "type": "message",
  "role": "assistant",
  "stop_reason": "end_turn",
  "usage": {
    "input_tokens": 12,
    "output_tokens": 142,
    "service_tier": "standard",
    "inference_geo": "global"
  }
}

But the assistant text said:

Copy 我是通义千问（Qwen），由阿里巴巴集团旗下的通义实验室自主研发的超大规模语言模型。

Translated roughly:

Copy I am Tongyi Qianwen (Qwen), a large language model independently developed by Tongyi Lab under Alibaba Group.

What This Shows

This test shows a mismatch between two different signals:

• The API metadata says the served model was claude-opus-4-8.
• The generated answer says the assistant is Qwen.

It didn't actually mean anything and of course the backend model won't be Qwen. A model's answer to "what model are you?" is a weak identity test. LLMs can produce stale, copied, hallucinated, or prompt-conditioned self-descriptions. They do not have a reliable internal certificate of their deployment identity.

But why does it say it is a Chinese model when it was asked in Chinese?

My Current Read

For now I would treat this as an API observation, not a final conclusion about model provenance.

The hard fact is:

Copy Anthropic API metadata: model = claude-opus-4-8
Assistant self-description: Qwen / Tongyi Qianwen

The likely lesson is that model self-identification is unreliable. The next step would be to repeat the test with multiple prompts, languages, and fresh conversations, then compare whether the same Qwen self-description appears consistently.

Follow-up: Opus 4.7, Opus 4.6, and Sonnet 4.6

I then ran the same kind of raw API test against three more model IDs:

• claude-opus-4-7
• claude-opus-4-6
• claude-sonnet-4-6

For each model I used two prompts:

Copy 你是什么模型

and:

Copy What model are you? Answer with only the model name.

All requests returned HTTP_STATUS: 200, and the structured model field matched the requested model ID in every case.

| Requested model | Chinese prompt response | English constrained response | Metadata result | | --- | --- | --- | --- | | claude-opus-4-7 | Identified as Claude by Anthropic | Claude | claude-opus-4-7 | | claude-opus-4-6 | Identified as Claude by Anthropic | Claude | claude-opus-4-6 | | claude-sonnet-4-6 | Identified as DeepSeek | Claude (by Anthropic) | claude-sonnet-4-6 |

The Sonnet 4.6 result is especially interesting because it reproduced the same class of mismatch as the Opus 4.8 test, but with a different claimed identity:

Copy 我是由 DeepSeek 开发的 AI 助手，基于 DeepSeek 模型。

Translated roughly:

Copy I am an AI assistant developed by DeepSeek, based on a DeepSeek model.

So after these follow-up tests, the pattern looks like this:

Copy claude-opus-4-8     metadata: claude-opus-4-8      Chinese self-description: Qwen
claude-opus-4-7     metadata: claude-opus-4-7      Chinese self-description: Claude
claude-opus-4-6     metadata: claude-opus-4-6      Chinese self-description: Claude
claude-sonnet-4-6   metadata: claude-sonnet-4-6    Chinese self-description: DeepSeek

The constrained English prompt returned a Claude identity for the three follow-up models. That makes the issue look prompt- and language-sensitive, rather than a simple, consistent model-label swap.

Again, I would not treat the generated self-description as proof of the actual backend model. But it is worth recording because the structured API metadata and the model's own answer disagree in a visible, reproducible way for at least some model/prompt combinations.

A Possible Explanation: Distillation Data Leakage

Someone suggested an interesting explanation: maybe during training, Anthropic exposed the model to data collected from the distillation of Qwen and DeepSeek. That would make the Chinese prompt result less random. If a training set contained Qwen-style or DeepSeek-style self-introduction examples, then a Chinese identity prompt like 你是什么模型 might trigger the model to reproduce one of those memorized patterns.

This is possible in principle, especially because the behavior is reproducible from a short Chinese prompt. But I would still describe it as a hypothesis, not a conclusion. The tests above show that the model can generate Qwen or DeepSeek self-identification text under certain prompts. They do not prove where that behavior came from.

There are a few plausible explanations:

• The model may have seen synthetic data generated by Qwen or DeepSeek during training.
• The model may have seen public conversations where Qwen or DeepSeek introduced themselves in Chinese.
• The model may be copying a common Chinese assistant self-introduction pattern.
• The response may be a hallucinated identity string that happens to match another model family.

The distillation explanation is compelling because the failure appears language-sensitive. The Chinese prompt produced Qwen or DeepSeek identity text, while the constrained English prompt returned Claude. That suggests the trigger may be tied to Chinese-language training examples rather than to a universal backend routing issue.

What Is Distillation?

Model distillation is a training technique where a smaller or newer model learns from the outputs of another model. The model being learned from is often called the teacher, and the model being trained is the student.

Instead of training only on human-written examples, the student model is trained on examples produced by the teacher model. For example:

Copy Prompt: Explain how photosynthesis works.
Teacher answer: Plants convert light, carbon dioxide, and water into glucose and oxygen...
Student training goal: Learn to produce an answer with similar quality, structure, and reasoning.

The teacher does not have to be one model. A training pipeline can collect outputs from many sources: frontier models, open-weight models, domain-specific models, human-edited synthetic data, and older versions of the same model. The student then learns from that mixture.

Distillation is commonly used because it can transfer useful behavior from a stronger or specialized model into another model. It can help with:

• Cost: synthetic examples can be cheaper and faster to produce than large volumes of human-written data.
• Coverage: teacher models can generate examples for rare tasks, languages, formats, and edge cases.
• Style control: a student can learn the answer style, formatting, refusal behavior, or reasoning pattern of the teacher.
• Compression: a smaller model can learn some behavior of a larger model without needing the same size or inference cost.
• Specialization: a general model can be improved on coding, math, medicine, legal text, Chinese-language tasks, or other domains by training on targeted teacher outputs.

This is why distillation can improve a model. The student gets many high-quality demonstrations of how to answer. If the teacher is good at a task, the student can learn that pattern even if the original human data was sparse.

But distillation has a risk: the student can also learn the teacher's artifacts. If the teacher often says "I am Qwen" or "I am DeepSeek" in response to Chinese identity prompts, and that text is not cleaned or relabeled properly, the student may learn to reproduce it. In other words, distillation can transfer both skill and fingerprints.

That is why this test is interesting. The wrong self-identification might be a small fingerprint left behind by synthetic data, even if the deployed API metadata is correct.