U.S. and China Reach Near-Parity in AI Model Performance, Report Finds
Listen to the full version
The gap in performance between top-tier artificial intelligence models developed by the U.S. and China has largely closed, marking a new phase in the global AI race — one characterized by diminishing technological dividends and highly concentrated resources, according to a recently released report.
Stanford University’s Institute for Human-Centered Artificial Intelligence published its 2026 Artificial Intelligence Index Report, which highlights that AI is making breakthroughs in areas such as scientific reasoning and coding, nearing human benchmarks. The focus of AI development has shifted; no longer are model performance scores the primary issue, but rather the massive expansion of infrastructure, the structural reshaping of the labor market, and increasingly urgent environmental and safety governance challenges.
Unlock exclusive discounts with a Caixin group subscription — ideal for teams and organizations.
Subscribe to both Caixin Global and The Wall Street Journal — for the price of one.
- RELATED
- DIGEST HUB
Digest Hub
Back
- US-China top AI models near parity (2.7% gap); exceed humans in science/coding (60-100% baselines), but lag in real-world tasks (e.g., 12% robot success).
- US leads models (50 vs. 30), investments ($285.9B private); China tops publications (41% top papers), robots.
- AI investment $581.7B; productivity up 14-50%, entry jobs down 20%; emissions high (Grok4: 72,816t CO2); incidents 362.
1. The performance gap between top U.S. and Chinese AI models has largely closed, entering a phase of diminishing technological returns and resource concentration, per Stanford’s 2026 AI Index Report.[para. 1]
2. The report highlights AI breakthroughs in scientific reasoning and coding nearing human levels, shifting focus from model scores to infrastructure growth, labor market changes, and environmental/safety challenges.[para. 2]
3. Annual AI reports since 2017 show accelerated capabilities reaching wider demographics; industry leads with >90% of notable models, surpassing humans in doctoral science, multimodal reasoning, math, and coding (60% to nearly 100% of human baseline in one year).[para. 3]
4. Gaps persist: AI wins IMO gold but reads analog clocks at 50.1% accuracy (vs. human 90.1%); robots hit 89.4% in simulations but 12% in real households.[para. 4]
5. In science/medicine, specialized small models outperform massive LLMs in biology/genomics; AI aids healthcare but 50% of studies use exam questions, only 5% real patient data.[para. 5]
6. Since 2025, U.S.-China models trade leaderboard tops: DeepSeek-R1 at 1,400 points trailed U.S. o1 by 0.4% (Feb 2025); by Mar 2026, Claude Opus 4.6 at 1,503 led Dola-Seed-2.0 by 2.7%.[para. 6]
7. U.S. leads in top models (50 vs. China’s 30 in 2025) and high-impact patents; China excels in publications (41% of top 100 cited papers in 2024, up from 33% in 2021), citations, patents granted, robot installs; competition shifts to costs, reliability, specialization.[para. 7]
8. U.S. has 5,427 AI data centers (>10x any other nation), but advanced chips depend entirely on Taiwan factories, exposing supply chain risks.[para. 8]
9. U.S. appeal to global AI talent wanes: researcher/developer relocations down 89% since 2017, 80% drop in 2025 alone, lowest rate in over a decade.[para. 9]
10. Global corporate AI investment doubled to $581.7B in 2025, generative AI at $170.9B (+200%); U.S. private $285.9B (23x China’s), China gov funds $184B (2000-2023); genAI adoption 53% population-level (Singapore 61%), faster than PCs/internet.[para. 10]
11. Productivity rises 14-50% in customer support, dev, marketing, but erodes entry-level jobs; U.S. 22-25yo developers down ~20% from 2022 peak; 1/3 organizations anticipate AI-driven layoffs.[para. 11]
12. AI’s environmental toll grows: Grok4 training emitted 72,816 metric tons CO2e; data centers at 29.6GW (NY state peak equivalent); GPT-4o inferencing water > needs of 12M people.[para. 12]
13. AI incidents rose to 362 in 2025 (from 233 in 2024); safety/privacy gains trade off against accuracy/fairness; frontier models least transparent, hiding training code, params, datasets.[para. 13]
- PODCAST
- MOST POPULAR
Sign in with Google
Sign in with Facebook
Sign in with 财新