Open-weight models versus frontier AI

Frontier AI labs (OpenAI, Anthropic, DeepMind), face an economic paradox: despite technological progress, their business model is unsustainable. Training costs grow exponentially, while capability prices fall ~10x annually due to efficiency gains and rapid open-weight distillation. This leaves only a 6-12 month exclusivity window to recoup massive fixed costs. To survive, token volume must grow ~30x yearly, yet better models burn fewer tokens per task, undermining revenue. Resembling unpatented pharmaceuticals, labs must abandon per-token pricing for outcome-based models or internal use to avoid the fate of the 1990s fibre-optic bust.

Based on research from Epoch AI, every model at the true frontier of AI capabilities since 2023 has been developed in the United States. Chinese frontier models have consistently lagged behind the US baseline by an average of seven months (ranging between 4 to 14 months).

The chart below visualises the growth trajectory of the Epoch Capabilities Index (ECI) for the top-performing US and Chinese models from January 2023 to mid-2026. Prior to April 2024, the ECI grew at roughly 8.3 points per year; following the rise of advanced reasoning models (like OpenAI’s o-series), progress accelerated to 15.5 points per year.

Key Timeline Dynamics

• The 2023–2024 Baseline: US models held a comfortable lead following the launch of GPT-4. Chinese providers spent this period building out foundational ecosystems (led by Alibaba’s Qwen and Baidu), tracking roughly 10–14 months behind the absolute US frontier.
• The April 2024 Inflection Point: AI progress globally experienced an inflection point around April 2024. The introduction of reinforcement learning and native reasoning paths nearly doubled the slope of capability gains from ~8 points/year to over 15 points/year.
• The 2025–2026 Open-Weight Catch-Up: By late 2025 and into 2026, Chinese labs narrowed the gap significantly. The release of highly efficient open-weight models like DeepSeek-R1 and Moonshot Kimi K3 brought China’s top tier within just 4 to 7 months of leading US closed-source systems, compressing the historical capability lag to its lowest margin yet.

Source:

Luke Emberson (2026), “Chinese AI models have lagged the US frontier by 7 months on average since 2023”. Published online at epoch.ai. Retrieved from ‘https://epoch.ai/data-insights/us-vs-china-eci‘ [online resource]. Accessed 9 Sept 2026.