Since DeepSeek R1 came out, many people have felt that Chinese models are catching up to the US frontier. Using a new benchmark, Frontier Clock, I find the opposite: while noisy, the US-China gap has grown by around 2.4 months per year.
Inspired by Epoch AI's “Chinese AI models have lagged the US frontier by 7 months on average since 2023” and Parv Mahajan's “The US-China AI gap is very uncertain,” I built an AI benchmark that measures the gaps between the US, China, non-China open weights, and individual labs. Instead of giving each model a traditional benchmark score, Frontier Clock assigns a “frontier-equivalent date”: the date when the world’s best publicly released model first had the same capability level as that model. It uses Epoch AI’s ECI benchmark data. The “Method” section in frontierclock.org discusses how the data is used.
This graph from the website plots each model’s frontier-equivalent date on the y-axis and release date on the x-axis:
Frontier Clock classifies models into three groups: “US Closed,” “China,” and “Open weights (non-China).” For each group, it displays the “gap to the frontier,” how many months each group’s best model is behind the world's best publicly released model, over time, from the ECI data.
The faint blue horizontal line over the X-axis represents the “US Closed” group. Closed-weight US models have always been at the frontier, so the value is always 0 months. The red line is China, and the teal line is non-Chinese open-weight models.
The red Chinese-model line is noisy, but the trendline shows it is increasing over time by 2.4 months per year. There is no sign that China is catching up; if anything, the US-China gap is growing. This trend has been discussed before.
How confident are we about this? Frontier Clock estimates uncertainty by randomly resampling across Epoch’s 47 benchmarks and fitting 90% confidence intervals. These intervals only capture uncertainty from which benchmarks are selected, so the true uncertainty is probably larger. Details are in the website’s “Method” section. Here is the same graph with confidence intervals:
The 90% intervals for China’s slope go from -2.0 to +3.3 months per year, dipping into negative values. Does this mean China might actually be catching up? Probably not. The intervals are skewed by massive uncertainty from September 2024 to January 2025, shown by the tall red bands on the left of the graph. From 2023 to late 2024, the frontier moved slowly; a small shift in ECI points results in a large shift in frontier-equivalent date, creating this huge uncertainty. I take four measurements that avoid this problem, and each results in nonnegative intervals:
Starting in January 2025 to skip the most uncertain period: +2.6 months/year (+0.8 to +3.9)
Starting in April 2025 to skip the low point following DeepSeek R1’s shock release: +1.7 months/year (0.0 to +3.9)
Excluding models with fewer than 5 benchmark values (instead of the site’s default of 2): +2.4 months/year (+0.1 to +2.8)
Measuring the gap in ECI points: +4.3 ECI points/year (+1.7 to +6.7)
China’s gap looks roughly flat from January 2026 onwards. This does not mean the capabilities gap is also flat. AI progress seems to be accelerating. A month behind today is a bigger gap in capabilities than a month behind was a year ago. The real capabilities gap between the US and China may be increasing faster than the gap in months. If so, the graph underestimates how fast the US lead is growing.
US closed frontier labs sometimes delay model releases and keep models internally. For example, Mythos was reportedly available internally as early as February 2026, even though Fable was only released publicly in June. OpenAI’s o3 model was previewed in December 2024 and released in April 2025. Chinese labs probably don’t hold models internally for nearly as long. Chinese models are less capable and less likely to cause safety-focused delays compared to Mythos and Astra. A more accurate reading of China’s gap may be to add a few extra months. This is discussed here.
The rest of this post looks at why China is not catching up and what that implies. My arguments are somewhat speculative and have been made elsewhere, but I hope discussing them here is useful.
Why does China remain behind?
China has less compute. Export controls limit what GPUs they can buy from the US, and Chinese chipmakers are not as advanced as the main western ones. Production for Chinese AI chips ramped up far later than Nvidia’s and Google’s. They have had less time to streamline their processes and iron out the kinks. Altogether, China is estimated to have one seventh the AI-relevant compute of the US.
US labs may have cracked down on distillation attempts. There is speculation and some evidence that some Chinese models are distilled from US models. Anthropic reports that DeepSeek, Moonshot AI, and MiniMax have attempted to distill Claude. OpenAI also reports attempted distillation. Recently, US labs have made distilling harder, potentially slowing Chinese AI progress.
Frontier US labs may be using their models internally to accelerate R&D.Anthropic and OpenAI report AI-assisted speedups. Chinese labs may not be able to use these models. Some model providers have restrictions on other labs using their models (e.g. section D4 of Anthropic’s Commercial Terms of Service), and frontier labs are likely using internal, unreleased models that few people outside the labs can access.
The above reasons are not exhaustive.
The US can afford to slow down.
If the US-China gap is increasing, then the US can slow development to match China’s current pace. This would keep the US steadily ahead, but might not slow things enough to address alignment and other safety risks in time. Slowing even further risks China catching up, but the US could still reach ASI first if the extra slowdown is limited; how limited depends on how far away ASI is. The slowdown can be temporary, ending if researchers and auditors gain sufficient confidence in alignment and safety techniques. A slowdown may also indirectly slow China by preventing them from distilling new models.
The most extreme version of a slowdown is a complete pause. Given the several-month lead, a temporary pause seems viable but may put the US in a precarious situation since the gap is volatile. However, the US could slow down or pause more safely if China agrees to do the same. I think this is plausible; being in the lead gives the US extra leverage in negotiations. Also, understanding that they are probably fighting a losing battle, China may be willing to cooperate with the US. If China thinks they will fall further behind, slowing down costs them relatively little, and a mutual slowdown also hinders the US, which China wants.
Trust is an issue, but if incentivized to seek a mutual slowdown, China may be willing to take explicit measures to gain the trust of the US (e.g. by increasing transparency in model training runs). Finally, Chinese officials have expressed worries about traditional AI safety risks and could want a slowdown regardless of which country is ahead.
Final notes
Frontier Clock has many caveats. A few taken from the website:
Model development can be very unpredictable. Measurements regarding how many months a certain group is behind reflect which benchmarks are included and may not be accurate. The actual gap may fall outside the 90% confidence intervals.
Raw benchmark scores are sourced from Epoch AI. Thus, Frontier Clock inherits any possible benchmark issues from Epoch such as potential contamination.
Only models at least as capable as GPT-4 (Mar 2023) are shown.
There are probably more I did not think of. These somewhat weaken the results, but I think the main implications still apply.
Frontier Clock is a new project, and I plan to continue updating it. I would love to hear your thoughts and feedback! Feel free to comment or reach out. Thanks for reading.
Since DeepSeek R1 came out, many people have felt that Chinese models are catching up to the US frontier. Using a new benchmark, Frontier Clock, I find the opposite: while noisy, the US-China gap has grown by around 2.4 months per year.
Inspired by Epoch AI's “Chinese AI models have lagged the US frontier by 7 months on average since 2023” and Parv Mahajan's “The US-China AI gap is very uncertain,” I built an AI benchmark that measures the gaps between the US, China, non-China open weights, and individual labs. Instead of giving each model a traditional benchmark score, Frontier Clock assigns a “frontier-equivalent date”: the date when the world’s best publicly released model first had the same capability level as that model. It uses Epoch AI’s ECI benchmark data. The “Method” section in frontierclock.org discusses how the data is used.
This graph from the website plots each model’s frontier-equivalent date on the y-axis and release date on the x-axis:
Frontier Clock classifies models into three groups: “US Closed,” “China,” and “Open weights (non-China).” For each group, it displays the “gap to the frontier,” how many months each group’s best model is behind the world's best publicly released model, over time, from the ECI data.
The faint blue horizontal line over the X-axis represents the “US Closed” group. Closed-weight US models have always been at the frontier, so the value is always 0 months. The red line is China, and the teal line is non-Chinese open-weight models.
The red Chinese-model line is noisy, but the trendline shows it is increasing over time by 2.4 months per year. There is no sign that China is catching up; if anything, the US-China gap is growing. This trend has been discussed before.
How confident are we about this? Frontier Clock estimates uncertainty by randomly resampling across Epoch’s 47 benchmarks and fitting 90% confidence intervals. These intervals only capture uncertainty from which benchmarks are selected, so the true uncertainty is probably larger. Details are in the website’s “Method” section. Here is the same graph with confidence intervals:
The 90% intervals for China’s slope go from -2.0 to +3.3 months per year, dipping into negative values. Does this mean China might actually be catching up? Probably not. The intervals are skewed by massive uncertainty from September 2024 to January 2025, shown by the tall red bands on the left of the graph. From 2023 to late 2024, the frontier moved slowly; a small shift in ECI points results in a large shift in frontier-equivalent date, creating this huge uncertainty. I take four measurements that avoid this problem, and each results in nonnegative intervals:
China’s gap looks roughly flat from January 2026 onwards. This does not mean the capabilities gap is also flat. AI progress seems to be accelerating. A month behind today is a bigger gap in capabilities than a month behind was a year ago. The real capabilities gap between the US and China may be increasing faster than the gap in months. If so, the graph underestimates how fast the US lead is growing.
US closed frontier labs sometimes delay model releases and keep models internally. For example, Mythos was reportedly available internally as early as February 2026, even though Fable was only released publicly in June. OpenAI’s o3 model was previewed in December 2024 and released in April 2025. Chinese labs probably don’t hold models internally for nearly as long. Chinese models are less capable and less likely to cause safety-focused delays compared to Mythos and Astra. A more accurate reading of China’s gap may be to add a few extra months. This is discussed here.
The rest of this post looks at why China is not catching up and what that implies. My arguments are somewhat speculative and have been made elsewhere, but I hope discussing them here is useful.
Why does China remain behind?
China has less compute. Export controls limit what GPUs they can buy from the US, and Chinese chipmakers are not as advanced as the main western ones. Production for Chinese AI chips ramped up far later than Nvidia’s and Google’s. They have had less time to streamline their processes and iron out the kinks. Altogether, China is estimated to have one seventh the AI-relevant compute of the US.
US labs may have cracked down on distillation attempts. There is speculation and some evidence that some Chinese models are distilled from US models. Anthropic reports that DeepSeek, Moonshot AI, and MiniMax have attempted to distill Claude. OpenAI also reports attempted distillation. Recently, US labs have made distilling harder, potentially slowing Chinese AI progress.
Frontier US labs may be using their models internally to accelerate R&D. Anthropic and OpenAI report AI-assisted speedups. Chinese labs may not be able to use these models. Some model providers have restrictions on other labs using their models (e.g. section D4 of Anthropic’s Commercial Terms of Service), and frontier labs are likely using internal, unreleased models that few people outside the labs can access.
The above reasons are not exhaustive.
The US can afford to slow down.
If the US-China gap is increasing, then the US can slow development to match China’s current pace. This would keep the US steadily ahead, but might not slow things enough to address alignment and other safety risks in time. Slowing even further risks China catching up, but the US could still reach ASI first if the extra slowdown is limited; how limited depends on how far away ASI is. The slowdown can be temporary, ending if researchers and auditors gain sufficient confidence in alignment and safety techniques. A slowdown may also indirectly slow China by preventing them from distilling new models.
The most extreme version of a slowdown is a complete pause. Given the several-month lead, a temporary pause seems viable but may put the US in a precarious situation since the gap is volatile. However, the US could slow down or pause more safely if China agrees to do the same. I think this is plausible; being in the lead gives the US extra leverage in negotiations. Also, understanding that they are probably fighting a losing battle, China may be willing to cooperate with the US. If China thinks they will fall further behind, slowing down costs them relatively little, and a mutual slowdown also hinders the US, which China wants.
Trust is an issue, but if incentivized to seek a mutual slowdown, China may be willing to take explicit measures to gain the trust of the US (e.g. by increasing transparency in model training runs). Finally, Chinese officials have expressed worries about traditional AI safety risks and could want a slowdown regardless of which country is ahead.
Final notes
Frontier Clock has many caveats. A few taken from the website:
There are probably more I did not think of. These somewhat weaken the results, but I think the main implications still apply.
Frontier Clock is a new project, and I plan to continue updating it. I would love to hear your thoughts and feedback! Feel free to comment or reach out. Thanks for reading.