Is China Catching Up to the US in AI?

_{^{Guest contributor: Anastasios Angelopoulos is an AI researcher, entrepreneur, and co-founder and CEO of Arena, advancing real-world AI evaluation, reliability, and machine learning research.}}

The question of whether China is catching up to the US is being argued both ways - and the data can be sliced to support either side.

Look at the data one question at a time, though, and a clearer picture emerges. The US still leads at the frontier. China has the deeper open-model bench, and it is increasingly competitive on cost. Who’s ahead depends on which race you’re watching.

At the very top of the leaderboards, the answer is still the US. Anthropic and OpenAI are releasing models relentlessly and are now significantly ahead on many benchmarks, including Agent Mode, which measures real-world performance in agent tasks (17.9% net improvement for Anthropic’s Opus 5.5 Max vs 3.3% for the highest-ranked Chinese model, DeepSeek V4.1 Flash Max).

That lead is real, but it is not safe. In July of 2026, we saw Kimi K3 overtake Fable 5 on the Code Arena; this was the first time that a Chinese lab has taken the #1 spot on the Arena AI Leaderboard. Today, Fable, Astra and Opus are well ahead of their Chinese rivals, but historical data indicates that the gap can close in an instant.

In the US, the competition is all fueled by release speed. OpenAI and Anthropic are neck and neck as their flagship releases have consistently landed two months apart all year long.

Stepping back, the larger story then is about who else is starting to keep up with the two leaders. SpaceXAI changed its strategy with the acquisition of Cursor and has now made a significant jump, especially in coding capability. This suggests that the company’s acquisition gave them the data to build a better model faster in order to compete at the US frontier.

The trend holds with Meta, as they also bet on speed. It shipped Muse Spark 1.1 through 1.3 in under two months time in an effort to stay at pace, and it has paid off. Muse Spark 1.3 (Max) has jumped over 20+ places above its 1.1 predecessor on the Agent Arena leaderboard.

Meanwhile, Google shows a different story. Its flagship models have moved more slowly. Gemini 4 Argon High currently tops the Text Arena leaderboard and has placed in the top five for Agent, but lags behind in coding, where the majority of the state-of-the-art models are focused. 

This level of competitive pacing doesn’t mean that the US is destined to win, but it just makes it all that much harder for anyone to catch up.

We have seen a huge shift in capabilities with proprietary models with Mythos/Fable, followed up by Astra and Opus on Agent Mode. Though we don’t know the details of the architecture, given the price per token differences (Fable 5.1 and Astra: 10/50 per million input/output tokens, vs 4/20 for Opus 5.5 and GPT-5.6 Sol), it is possible that Fable and Astra are models that have been scaled up in size. 

Chinese labs are scaling up too, from 1T total parameters for Kimi K2.5 to 2.8T for Kimi K3, while Qwen3.8 has 2.4T total parameters. But it is not clear that the scale-up has contributed enough to performance, given that they are still behind (Qwen3.8 Max 0902 is 10th on Code Arena and Kimi K3 Max is 14th).

The US labs also have far more resources to keep pushing. Revenue of Anthropic and OpenAI is now far above that of the Chinese counterparts shown here.

Table 1: Reported annualised revenue estimates

The US has also been undertaking a significant level of investment in AI, with Stargate plans for over $400B over three years, and with projects coming online on different schedules, there is still significant computing capacity yet to come.

China’s labs have so far managed to repeatedly respond and catch up to their US counterparts, and their compute investments have also been significant; e.g., Alibaba is planning at least RMB380bn (US$53B) for AI and cloud infrastructure over three years, announced in February 2025. Whether China can gather enough compute to keep pace is the open question.

For now, the US leads at the frontier and has the revenue and compute to keep pushing. But China has shown it can close the gap fast.

Here, the answer is China. The level of competition within China itself is enormous. DeepSeek and Alibaba were among the labs with notable models in both 2024 and 2025. But since then, we have seen very credible models come out of labs from Moonshot AI, Z.ai, MiniMax, Tencent, Xiaomi and StepFun.

Compared to the US, China actually has more labs operating at the frontier - if we look at the top 15 labs on Code Arena, we see nine labs from China, five from the US and one from France.  

One obvious area where China dominates the US is open-weight models. If we look at the current open-weight leaderboards - Agent Mode, Code Arena, or Text - Chinese models take up the top five positions. 

Open-source is very much part of the core China strategy. This contrasts with the US approach, where leading labs have focused on proprietary models and have specifically dedicated open model lines, such as Gemma by Google or GPT-OSS (though OpenAI’s general-purpose GPT-OSS family is 14 months old). In recent months, new labs and institutes such as Thinking Machines, Ai2, Reflection and Poolside are now also looking to challenge the market with their open models. The reason is simple: businesses prefer to use open-source models so they can control the supply chain of their AI products.

Nvidia is poised to benefit most from this, as AI adoption drives demand for chips. This is why they have been increasingly investing in the US open-source ecosystem through the Nemotron series of models, reportedly striking an investment and licensing deal with Poolside, and reportedly holding acquisition or further-investment talks with Reflection. They have also agreed to acquire HuggingFace. It seems like Nvidia is treating availability of competitive open models as a strategic imperative, though Chinese models still lead the current open-weight Arena leaderboards.

The challenge with open models growing in size is that the released weights for the 2.8T Kimi K3 model alone are 1.56TB, which require, by capacity alone, 20 H100 80GB GPUs or five GB200s, before cache and runtime overhead, meaning that in-house hosting is demanding for households and businesses. This could mean that many are relying on a neo-cloud to provide the inference. 

This also gives us a hint at a business model that companies like Moonshot AI might be exploring with open models. Their open license has a clause that 3rd-party hosting companies running a “Model as a Service” business with group revenue over $20m in any consecutive 12 months need to have a separate agreement with Moonshot AI. While details are not publicly known, it is not hard to imagine a revenue share model. Meaning that users who use the Kimi K3 model could be paying Moonshot either directly or indirectly, depending on those terms.

China has the deeper bench: more labs near the frontier and the top open-weight models on every current Arena leaderboard. The US open-model response is real, with Nvidia leading the charge, but it has not yet produced a model that tops the open-weight rankings.

This is where China’s open-model strategy turns into pressure on the US labs. But price per token only tells part of the story.

Some labs have discovered that scoring well on benchmarks could be achieved by using more computation to solve the task. While this could give a chance for lower performance models to compete, more generated tokens can create a detrimental experience for users since they would need to pay for the tokens and could experience increased latency. 

We have seen in the Agent Mode data, which measures performance based on real-world tasks from tens of millions of global users, that Astra Max uses 31,739 median output tokens per task and strong models like Fable 5.1 Max have equally been able to demonstrate very strong performance with fewer tokens than Opus 5.5 Max (58,000 vs 113,000). 

However, there are no Chinese models on the current token efficiency frontier among the models shown here. Kimi K3 Max uses similar median output tokens as Astra Max (31,000 vs 32,000), though its score is lower. This might point to one way performance on static benchmarks might differ from the real-world user experience.

Token efficiency is only part of the equation; what we are ultimately interested in is performance vs. cost per task. Agent Arena’s Pareto curve is the best way to see this - since we can track the observed real-world performance combined with median real-world costs per segmented task, taking into account important hidden parameters like caching and token consumption. 

Once we plot the curve, we can see a near-perfect split - on the US-China median-cost Pareto frontier, both points above $0.50 per task are American, and most positive-score frontier points below $0.50 are Chinese, with the exception of GPT-6-Luna.

Open models have an advantage as they benefit from industry-wide innovation in inference efficiency improvements, meaning that the cost can go down after release, potentially pressuring the margins of the American labs. We have seen significant cost reductions from OpenAI and Anthropic (80% for GPT-5.6 Luna input/output prices in July and 50% for Sonnet 5.5 cache reads in October), which could mean that open models have some cost-pressure effects on the labs. 

There is also a solid case for ‘good enough’ models. If in ~6-8 months Chinese labs have Opus 5.5-level models, some share of the market may be satisfied with that level of performance and focus on cost instead. And with no US open model ahead of the Chinese open models on the current open-weight Arena leaderboards, there is a significant chance that the low/mid end of the market could belong to Chinese open models.

US models are more efficient at the top end, but China increasingly sets the price for everything below it.

American labs have the best models, the most revenue and the deepest compute pipeline. Chinese labs have the deeper open-model bench and the lowest cost per task, and they have shown they can close a frontier gap fast.

That pressure is already showing up in US pricing. Three questions will decide how far it goes: Will ‘good enough’ models become the norm? Will China be able to gather enough compute to compete? Will US labs mount a significant open-source response?

Whatever the answers, the market is better for it. With this many labs across the US and China competing for the frontier, no single lab can easily dominate, which should lead to both rapid capability improvements and cost reductions - which is great for consumers across the world.