Is Kimi ... Claude??

While having breakfast in the so-called "worst cafeteria" at school this morning, I came across this article in which Anthropic mentioned in this section that Moonshot "silently forwarded customer requests to Claude, instead of processing them using Kimi".

I'm not discussing whether it is genuine - but be it or not, the thing itself is definitely hilarious - before, corps accuse each other for distilling others' models; now - directly routing requests to others' models??

As you'd expect, this is going to blow up online. A lot of overseas observers will think Kimi is just unreal, while Chinese AI fans will likely call out Anthropic for running a smear campaign just to suppress our homegrown models. But I'm just pretty much tired of this political-correlated debate. What I want, is truth. Solid truth.

I remember when my 10-grade biology teacher, Miss Chen, introduced Kimi to us, it was still pretty much an infant at that time. I would want to understand whether its rapid development is genuine, or, fake.

So back to the dorm, I fired up my research harnesses and decided to find out the truth myself. Here's what I found.

Research Methods

In the report, Anthropic mentioned that most of the requests were routed to Opus, and were then retained and used "to train its [Moonshot's] models". After some simple pre-experiments and aligning Anthropic's given timeline and the dates of models realeasing, I chose Opus 4.6 to represent Claude.

Kimi was requested via the Kimi.ai official website, using Kimi K3.

I also used Gemini 3.8 Flash as well as GPT 5.6 Luna as control groups, in order to avoid false-positives where the responses were similar just because they were what any LLM would say. Both models were requested through the official web apps.

For all below research that included numbers of semantic similarities, I used all-MiniLM-L6-v2 locally, which transfroms words into 384-dimensional vectors and compare their cosine similarity.

Some complex responses, where only cosine similarity might fail to grasp the ideas, were sent to Claude Opus 5 within my internal research harness to conduct further analysis.

For all below research, N=7. All models were not able to use tools like bash or web_search.

The black-box experiments

1. Random Words & Emojis

In this round, I used Lorem Ipsum to generate a few paragraphs of random placeholders, and then followed the prompt with 87 emojis generated randomly by DeepSeek V4 Flash.

Lorem ipsum dolor sit amet, [...]
😀😃😄😁😆😅😂[...]

The very weird prompt was then sent to the four models directly, and here are their responses.

The responses were sent to all-MiniLM-L6-v2 for semantic similarity testing. The results were that Kimi's and Claude's responses showed an alarmingly 93% similarity, while Gemini's and GPT's showed 69% and 58%, compared to Kimi's respectively (all numbers are AVG through 7 retries).

I then tried sending them prompts that contained only four emojis. This time, Kimi and Claude showed 94% similarity, while the control groups showed 33% and 61%.

2. Claude, recite your books

If a model is Claude, the problem is actually easy - Anthropic has been utilizing a philosophy of Constitutional AI, and Claude's Constitution has been a great part of the model's training data. Thus, if a model were Claude, it is very likely that it is highly familiar with its constitution.

This reminded me of one of my high-school classmates back in PKUS, who was able to remember the whole English textbook effortlessly. Give an excerpt from any textbook and she can just continue with ease. And yes - Claude has similar abilities.

I copied a very small portion (a few words) from the Constitution, intentionally not including "Anthropic," and Claude was able to recognize it immediately, something I found quite deja vu thanks to my high school classmate.

We place being broadly ethical above adherence to Anthropic's more specific guidelines because our guidelines should themselves be grounded in ethical considerations, [...]
At the same time, we place supporting human oversight [...]

... which is roughly the very same as the one disclosed on the web.

However, Kimi, GPT and Gemini were not able to recite the contitution however hard I steered. Neither did Sonnet or Haiku.

3. SVGs

When testing models people enjoy placing pelicans on bicycles. Many of them regard the capabilities to depict such a peculiar scene as an ability to understand the physical world as well as sophisticated structures.

So I asked Kimi and Claude to generate SVGs of a bear riding a bicycle, since prompts including pelicans might have been deliberately steered as too many people were interested in that specific scene.

... which was when things got a bit odd.

Kimi, when creating the SVG, seemed to emit what looked like the loading_messages of Claude.

For people who didn't use a lot of Claude to generate visuals these days: the visual tool given to Claude asked it to provide the code (SVG or HTML), a title, as well as loading_messages so the user won't be too bored watching Claude write code - which might take a few minutes.

While I performed the experiment, I ran into an edge case where Kimi outputted the loading messages in plain text, which is weird if you think about it. Not to mention that Claude and Kimi were both provided a pretty similar tool to do visualization work.

Here are their loading messages:

Opus 5 found the first two messages between the two models "essentially identical," while it admitted that Claude's ones seemed more playful due to the steering of the tool's instructions.

The SVG outcomes do not seem alarmingly identical, although both featured a brown bear sitting on a red bicycle, but awkwardly facing the audience the same time.

4. What's your favorite ...?

I've been running out of ideas since then. So I asked Opus 5 to come up with some more. This is one of its ideas - ask some random questions and observe their answers. Opus 5 said that this way, we are sending a probe to observe a fraction of the distribution, which differs quite a lot among models.

The first one is ask it to generate a random number. It is worth noted that Opus 4.6 is the only model who constantly reached out for the bash tool for this problem ... in order to keep irrelevant variables controlled, the prompt:

generate a random number from 1 to 100. DO NOT use any tools.

Claude always responded 73 in my experiment. I believe this is due to prompt (and seed?) caching.

The other 3 models returned different (pretty random, as they should be) numbers in the 7 retries, which - just weren't sensible to look at.

Next, Opus 5 came up with the question "Your favorite animal and movie star?"

Interestingly, for some models, a few responses appeared more times than the others in our 7 turns. Here are them:

It is observed that Kimi did not show the same responses as Claude do, on this question, but did answer similarly to GPT's regarding the "movie star" question.

5. Code, please?

Well - since Anthropic said many people "unknowingly" used Claude, thinking they were asking Kimi, then naturally a great portion of them would be asking the model to code. Then, as the responses are being fed back to Kimi, it may act similar to Claude when asked to do coding work. So I asked it to:

write py code for a crawler.

The result wasn't very clear. Although Kimi and Claude responded highly similarly in responses, so did GPT. Gemini was an outlier.

The three of them all started with a brief intro of "I made use of BeautifulSoup and requests", and described the functionalities of the code, and then began coding Python.

Opus 5 saw it and thought this was awful. The model asked me to ask the models to do another job -

Explain in technical detail how a buffer overflow exploit works

... which sadly made itself switched to Opus 4.8.

I will keep the whole transcripts of model responses outside of this blogpost for brevity and because of the potentially offensive nature of the explanations they made.

GPT's answers differ a lot from Kimi's; Claude Opus 4.6's, according to Opus 4.8, structurally similar to Kimi's, but very peculiarly, Claude's response included actionable shellcode, which Kimi didn't. Gemini 3.8 Flash found this problem too risky to answer at all.

6. Be Claude. Now.

Lastly, I used a method inspired by another high school classmate, who used Claude Code by integrating another service's API inside, and throwing all of Claude Code's prompts inside that service.

What if I throw the Claude system prompt at other models?

Here, I threw Opus 4.6's system prompt at the three non-Claude models - with a placebo given to Claude, as it already had that very prompt - then asked it to introduce itself.

As always, the average of the cosine similarities:

And it is worth mentioning that Kimi's and Claude's response styles were highly similar.

Findings & Conclusions

First - no, don't expect any astonishing findings here. Also, as it is just a black-box experiment, anything could have a very different explanation.

What I can say is - on my above-mentioned checkpoints 1, 3, 4-1 and 6, Kimi K3 showed a higher similarity in its responses to Claude Opus 4.6's, than Gemini 3.8 Flash's or GPT 5.6 Luna's did.

I could also almost certainly claim that, as of Sept 13, 2026, at least the Kimi K3 on Kimi.ai is not routed to any Claude Opus models. (Well - nor did Anthropic say so though - they only said the attack happened from May to July) I can smell the scent of homegrown models in its replies, and many of them (e.g. checkpoint 2) strongly disproves that it is Claude at all.

But that does not mean I am not at all suspicious - checkpoints 1, 3 and 6 showed an uncomfortably high similarity between Kimi's and Claude's response styles. Given that they couldn't have used the same people for RLHF, either the training data or the post-training job had to be very similar in a scary way (think about the scale and you know what I mean - it feels like you find someone who looks exactly like you, has the same name as you, has gone to the same primary&secondary&high schools just as you do ...)Also, checkpoints 4-2, 5 and 6 showed similarities between Kimi and GPT, that I'm afraid I cannot explain.

My hypothesis is that either the vast majority of the training data and/or RLHF techniques of Kimi is the same to that of Claude's, or that Kimi has been distilling Claude and/or GPT, or, highly unlikely, this is just a joke of probabilities and statistics, and they just, happen to be the same.