← Back to blogYouTube Video

Published November 16, 2026

"It's Not Even Close" - China vs. Opus 5.5

Can't play the video or having issues? Here's the direct link.

Summary

Two months after Opus 5.5 launched on September 22, no Chinese open-weight model has come close to it, and I think the crackdown on distillation (training a smaller model on a bigger model's outputs) is the reason. Chinese labs have always shipped cheaper versions of US frontier models rather than better ones, and a student can't surpass its teacher. Once Anthropic and OpenAI hid reasoning traces and added guardrails to stop distillation, releases like Kimi K3.1 (benchmaxed, weak on real tasks), Minimax M3.1 Flash and Ling 3.1 Flash fell flat. DeepSeek V4.1 Flash 1017 is the only one I use regularly, mainly when Anthropic blocks legitimate work such as reverse engineering my own hardware. My take is that I prefer open source and want it to catch up soon, but until then I'm stuck with Opus because the gap is that large.

Key takeaways

  • Distillation caps a model at its teacher's level: training a smaller model on a frontier model's outputs yields a cheaper version, never a better one.
  • Treat "90% as good as Opus at a fraction of the cost" claims with skepticism; cheap imitation isn't the same as frontier capability.
  • Test models on real-world tasks, not benchmarks: some releases, like Kimi K3.1, look good on leaderboards but struggle in actual work.
  • Keep an open-weight model as a fallback for work that closed models refuse; DeepSeek V4.1 Flash 1017 is my pick for tasks Opus blocks.
  • Models are tools, and I don't want an LLM deciding what's right or wrong for me; that's why I value open models that do the task even if they voice concerns.
  • For the hardest work, use the most capable model available, because settling for a weaker one means steering it far more than a strong model like Opus 5.5.

Description

Two months after Opus 5.5 came out, no Chinese open-weight model has gotten close to it, and I think the crackdown on distillation (training a smaller model on a bigger model's outputs) is the reason. I go through the recent releases like Kimi K3.1 and Minimax M3.1 Flash, and why DeepSeek V4.1 Flash 1017 is the only one I use regularly. I also explain why Anthropic blocking my requests keeps pushing me toward open models, and why I want open source to catch up soon.

Transcript

Auto-generated transcript
Chinese AI models suck right now. It's been months and China has not shipped a single open-weight decent model that can compete with the US frontier models. For context, Opus 5.5, the current leading model, came out on September 22. Two months have passed since then and nothing, not even open and not closed models have caught up to it. Like it's not even close. And two months is a big time in the AI world. Two months is like 10 years passing when it comes to the AI landscape. Now, it wasn't always like this, right? Several months ago, Chinese models were really only a couple steps behind, and they were shipping new models quickly. And now they've gone completely quiet. Now, they haven't gone completely dark, obviously. There have been some new open models released in the last two months, such as Kimi K3.1, DeepSeek V4.1 1017, some models from minimax and as well as a couple other labs that you probably haven't even heard of problem is all but one of them are completely garbage and horrible to use there is literally only one chinese model that is worth using right now and that i use on a regular basis and i'll show you which model that is but even that model is nothing compared to opus 5.5 and the other frontier models that the us labs are building now why is all of this happening and why can't china catch up with the US labs. Let's take a look at the market right now. Now as I mentioned before Opus 5.5 released on September 22. Since then two months have passed and no model not open or closed has caught up to it. It's still the best model on the market by far and it is so good that it's even beating Fable 5.2 which is Anthropic's own latest model. It beats that in some particular tasks. Now early in this race several months ago and ever since you know china started putting out its own models they've been catching up to the u.s models pretty fast they've always you know caught up to them but they've always also been a couple steps behind because if you look closely chinese models have always been pitched as a cheaper version of sonnet or opus or an open ai model right never a better version of the model. Like it's always this particular XYZ Chinese model that's supposed to be 90% as good as Opus at 100 the cost or something, right? It's never something that's better than Opus is what I'm trying to say. They're good at making cheap stuff. They're not good at making better stuff. Now the main technique that all of these Chinese labs have been using all this time to train and produce cheaper versions of these US frontier models is a little something called distillation and you need to understand what that means before we proceed with this video and so to put it in very simple terms distillation basically means training a smaller model on the outputs of a bigger frontier model For example if I wanted to distill a model, what I could do is I could send tasks or prompts to the latest frontier model, let's say something like Opus 5.5. And based on the results that it gives me, I would train a smaller model that I'm currently building, let's say Kimi K3.1 or the upcoming Coin4 release, right? So I tell Opus to go do some particular task for me. It does a bunch of thinking, it does a bunch of reasoning, it does a bunch of tool calls, and ultimately gives me the final response, right? I can then go and take all of that data and then train my own model on that, right? This is also the reason why Chinese labs have been so obsessed with reading the reasoning slash thinking traces of these US models and also why open AI and anthropic models don't actually show you the reasoning traces anymore, right? Like you can't see that thinking text of the model and see exactly what words the model is thinking, right? That's because they've just hidden it, right? Because we don't want to train these other models on that information, right? And of course, you can support whichever political side you want, but that is the reality of things, right? That's why we don't get to see the reasoning traces of all of these closed models anymore. So this is how Chinese models have been able to ship cheaper versions of all of these frontier US models all this time, right? And it is also the reason why Chinese models have never been able to catch up with US models. Because at least when it comes to AI models, the student cannot surpass the teacher, all right? If you distill a smaller model from the outputs of a bigger model, that smaller model is not going to beat the bigger model at any of those tasks. It is literally only designed to be a cheaper version of that bigger model that can do most of those tasks, but not as well as the bigger model, just cheaper than the bigger model. Now, US labs like Anthropic and OpenAI have been calling out these distillation attacks from China for a very long time. And you know, people have mixed opinions on that because after all, these US labs train their models on the entire free internet, and now they get to complain if someone else trains on their models. So, you know, they have mixed opinions on that. Not the point of this video. What is the point, though, is that Anthropic and OpenAI, but mostly Anthropic, has put even more guardrails into their models, especially when it comes to Chinese users. And the only reason they're doing this is to prevent distillation attacks. And they have been very, very successful with that. And now that these distillation attempts have been completely eliminated, China has struggled to put out a single decent open rate model And I think that tells you a lot about how they were trading their models before Because I know a lot of people who used to say that distillation attacks weren even a thing and I would encourage them to comment down below and tell me exactly why China is unable to produce a good model ever since we cracked down on the distillation attempts because I would love to see a good answer on that because right now it seems like the US labs were right China was distilling all the models and And ever since that stop, they've been unable to produce literally any decent good model. By the way, if this explanation made sense and you now have a better understanding of why Chinese models are falling off, consider liking and subscribing to the channel. Really helps me out as well. And you can get AI news like this without all of the hype. Now let's briefly talk about the recent Chinese releases that have come out and why they've been so disappointing. There was Kimi K3.1, very disappointing release. a Benchmaxed version of some OpenAI model. It struggles to do any kind of real-world tasks. It is literally only designed to look good in benchmarks. Minimax M3.1 Flash and Ling 3.1 Flash, both of them released this October. Not that good at all. I don't know a single person using those models. The one exception is DeepSeek V4.1 Flash 1017. It is the only Chinese model that I use on a regular basis that is actually very good but still it's nothing compared to Opus 5.5. It is literally only a model that I use anytime Opus blocks my requests because I'm doing some reverse engineering or I'm doing some other thing that Anthropic has decided that I can't or cannot do. And speaking of Anthropic, they are currently dominating hard. They released Fable 5.2 just a couple days ago. It is an extremely capable model and they've still got Fable 5.5 in their back pocket. They haven't released that yet. And even if they didn't, Opus 5.5 is such a capable model that it has been dominating the AI landscape for the last two months, which is again a huge amount of time in the AI world. We should have had like 20 different models being released by all these labs by now. And as if all of that was not enough on its own, Anthropic is also going to get hundreds of billions of dollars in new investment as soon as they launch their IPO, right because they're IPOing as a company now. Even OpenAI is struggling so hard to catch up right now. They don't have a single model that can compete with Opus 5.5 let alone all the other models that Entropic has in its back pocket. So if OpenAI is struggling to catch up I don't have much faith for Chinese labs either because the gap between Opus 5.5 and literally every other model in the entire AI market is extremely large. It's not even close. Now here's what I think about I like open source. I prefer open source models and I would like nothing more than for open source to catch up with Anthropic right now But the reality is Opus 5 is just way too good I can imagine not using this model Anytime I have to use a model that not Opus 5 it's just a horrible experience because I'm just used to giving like a one sentence prompt and having the model generate everything for me because Opus is that good. I don't enjoy steering all these smaller models anymore. But the problem is Anthropic's guardrails and model policing consistently block my requests. For example, it won't let me tinker with my own hardware, even though I own it, because that qualifies as reverse engineering. And Anthropic has decided in their infinite wisdom that that is somehow wrong for me to do. And this is why I love using DeepSeek V4-1-1017, because it will do the task that I gave it to do. It might tell me the right and wrongs of a particular task, but it's not going to prevent me from doing it. And that is really important because I don't want some LLM model to decide what's right and wrong for me. All right. I'm perfectly capable of doing that myself. I don't need some model to do that for me because at the end of the day, these things are just tools. All right. You don't expect a tool to decide whether you get to use it or not. And so I'm using DeepSeek probably every single day for tasks where the models from Andropic and OpenAI reject my requests completely. And so I really hope that the open source labs catch up over here because I don't want to use these models from Anthropic that constantly block my requests for completely legitimate work just because they feel like they have a better understanding of right and wrong than their users. I hate that kind of frame and I don't support it whatsoever. Problem is they have the best models right now by far. And so I just have to use them. I can't use a shitty version of Opus 5.5. It's just not that good. I have to use the most capable models. And currently, Anthropic is the only one that has them. That's just how things are. So to recap all of that, US models are currently leading the frontier, as always. Chinese labs all this time were relying on distillation attempts to create cheaper and less capable versions of the OpenAI and Anthropic models. And since the crackdown on distillation attempts, China has struggled to put out a single decent openweight model that can compete with US Labs with the sole exception of DeepSeek V4.1 Flash 1017. And I'm really hoping that Chinese and openweight labs catch up very soon because I don't want to depend on using closed models from companies that somehow think they're better qualified to decide what's right and wrong than their own users. So here's hoping open source catches up. That's all for this one. Let me know if you learned something new and whatever thoughts you have on this would love to read them in the comments below other than that like comment share subscribe do all of those wonderful things and i'll see you in the next one see ya

Share this article

Hassan Aziz pointing up

Dude, you’re at the bottom of my landing page.

Just book a call already if you’re that interested.

You scrolled all the way here.