2025 Stanford HAI Report: The gap between China and the US in AI models has narrowed to 0.3%, with inference costs plummeting by 280 times

The impact of artificial intelligence (AI) has never penetrated society as deeply as it does now. On April 8, Stanford University's Artificial Intelligence Institute (Stanford HAI) released the "2025 AI Index Report," which states that reasoning costs have plummeted by 280 times, the performance of small models has surged, and AI is becoming more efficient and accessible.

The report also points out that the quantity and quality of high-performance AI models in China are continuously improving, posing a challenge to the United States' leading position, and the performance gap between top models is narrowing.

The United States' previous advantage in model quality has disappeared. China is now the country with the most AI publications and patents, and the models it develops are now on par with American competitors in terms of performance.

"In 2023, in the large-scale multi-task language understanding test (MMLU), China's leading models lagged behind the top American models by nearly 20 percentage points. However, by the end of 2024, the U.S. lead shrank to 0.3 percentage points."

The latest blog post from Stanford HAI condenses the twelve highlights of the 2025 AI Index Report.

1. AI Performance Soars: Explosive Growth Under New Benchmarks

In 2023, researchers introduced new benchmark tests, such as MMMU, GPQA, and SWE-bench, to test the limits of advanced AI systems.

Just one year later, performance has dramatically improved: scores on MMMU, GPQA, and SWE-bench increased by 18.8, 48.9, and 67.3 percentage points, respectively. In addition to benchmark testing, AI systems have also made significant progress in generating high-quality videos, and in some cases, language model agents have even surpassed humans in programming tasks, despite limited time budgets.

2. AI Penetrates Daily Life: From the Laboratory to Reality

From healthcare to transportation, AI is rapidly moving from the laboratory to everyday life.

In 2023, the U.S. Food and Drug Administration (FDA) approved 223 AI medical devices, compared to only 6 in 2015.

On the roads, autonomous vehicles are no longer experiments: Waymo provides over 150,000 autonomous driving services weekly, and Baidu's Apollo Go autonomous taxi fleet now serves multiple cities in China.

3. Corporate Bets on AI: Investment and Application Soar

In 2024, private AI investment in the United States grew to $109.1 billion, 24 times that of the UK's $4.5 billion.

Generative AI performed particularly strongly, attracting $33.9 billion in private investment globally, an increase of 18.7% compared to 2023. The use of AI in business is also accelerating: 78% of organizations reported using AI in 2024, up from 55% the previous year.

An increasing number of studies confirm that AI enhances productivity and, in most cases, helps bridge the skills gap in the workforce.

4. The U.S. Still Leads in AI Models: But China is Closing the Gap

In 2024, U.S.-based institutions produced 40 noteworthy AI models, far exceeding Europe's 3.

While the U.S. maintains a numerical lead, Chinese models are rapidly closing the quality gap: The performance gap on major benchmark tests such as MMLU and HumanEval shrank from double digits in 2023 to nearly even in 2024.

Meanwhile, China continues to lead in AI publications and patents. At the same time, model development is becoming increasingly globalized, with projects from regions such as the Middle East, Latin America, and Southeast Asia also gaining attention.

5. Uneven Development of Responsible AI Ecosystem

AI-related incidents are sharply rising, but standardized RAI assessments remain rare among major industrial model developers.

However, new benchmark tests such as HELM Safety, AIR-Bench, and FACTS provide promising tools for assessing factuality and safety. Within companies, there remains a gap between recognizing RAI risks and taking meaningful action.

In contrast, governments are showing increasing urgency: In 2024, global cooperation on AI governance intensified, with organizations including the OECD, the European Union, the United Nations, and the African Union releasing frameworks focused on transparency, trustworthiness, and other core responsible AI principles.

6. Global Optimism for AI Rises: Regional Differences Remain Significant

In countries such as China (83%), Indonesia (80%), and Thailand (77%), the majority believe the benefits of AI products and services outweigh the harms. In contrast, places like Canada (40%), the United States (39%), and the Netherlands (36%) exhibit lower levels of optimism Nevertheless, sentiment is changing: since 2022, optimism has significantly increased in several previously skeptical countries, including Germany (+10%), France (+10%), Canada (+8%), the UK (+8%), and the US (+4%).

7. AI is becoming more efficient, economical, and accessible

Driven by increasingly powerful small models, the cost of system reasoning for tasks at the GPT-3.5 level has decreased by more than 280 times from November 2022 to October 2024.

On the hardware front, costs are decreasing by 30% annually, while energy efficiency is improving by 40% each year. Open-source models are also narrowing the gap with closed-source models, with performance differences shrinking from 8% to 1.7% in certain benchmark tests. These trends are rapidly lowering the barriers to advanced AI.

8. Governments are increasing investment in AI: Balancing regulation and investment

In 2024, US federal agencies introduced 59 AI-related regulations, more than double that of 2023, and issued by more than double the number of agencies. Since 2023, the legislation mentioning AI in 75 countries worldwide has increased by 21.3%, growing ninefold since 2016.

In addition to the growing attention, governments are making large-scale investments: Canada has committed to investing $2.4 billion, France has pledged €109 billion, India has promised $1.25 billion, and Saudi Arabia's "Beyond" project represents a $100 billion plan.

9. Expansion of AI and computer science education: Opportunities and challenges coexist

Currently, two-thirds of countries offer or plan to offer K-12 computer science education, double that of 2019, with Africa and Latin America making the most progress.

In the US, the number of graduates with a bachelor's degree in computer science has increased by 22% over the past decade. However, many African countries still struggle to access education due to gaps in basic infrastructure such as electricity. In the US, 81% of K-12 computer science teachers believe AI should be part of foundational computer science education, but less than half feel capable of teaching it.

10. The Industry Accelerates Development in the AI Field: Frontline Competition Becomes Increasingly Intense

In 2024, nearly 90% of renowned AI models come from the industry, up from 60% in 2023, while academia remains the primary source of highly cited research. The scale of models continues to grow rapidly, with training compute doubling every five months, datasets doubling every eight months, and power consumption increasing annually.

However, the performance gap is narrowing: The score difference between the top-ranked and tenth-ranked models has decreased from 11.9% to 5.4% within a year, and the gap between the top two is only 0.7%. Competition in the frontier field is becoming increasingly intense and crowded.

11. The Influence of AI in the Scientific Field is Recognized: Awarded Top Honors

The growing importance of AI is reflected in major scientific awards: two Nobel Prizes recognized work leading to deep learning (Physics) and its application in protein folding (Chemistry), while the Turing Award honored pioneering contributions to reinforcement learning.

12. Complex Reasoning Remains a Challenge: Limitations Still Exist

AI models excel at tasks such as solving International Mathematical Olympiad problems but still face challenges in complex reasoning benchmarks like PlanBench. They often fail to reliably solve logical tasks, even when provably correct solutions exist, limiting their effectiveness in high-risk environments where precision is crucial.