Originally published at news.skila.ai
China and the US are now neck-and-neck in AI model performance. That single data point, buried on page 47 of Stanford HAI's 2026 AI Index Report, rewrites the narrative the entire AI industry has been selling for three years.
Stanford released its ninth annual AI Index on April 13, 2026. The report covers AI research output, industry investment, workforce impact, public sentiment, and environmental cost across 300+ pages of data-driven analysis. Most coverage focused on the headline numbers. The real story is in what those numbers contradict.
Here are five rankings from the report that break assumptions you probably still hold.
1. China Erased the US Performance Lead
For years, the conventional wisdom was simple: America leads AI, China follows. The 2026 AI Index demolishes that framing.
US and Chinese models have traded places at the top of Arena community rankings multiple times since early 2025. As of March 2026, Anthropic's Claude Opus 4.6 leads globally, but the margin is razor-thin: 2.7% ahead of the best Chinese models.
China's strengths have shifted. The country now leads in AI patents, academic publications, and autonomous robotics deployment. China installed 295,000 industrial robots in 2024. Japan installed 44,500. The US? Just 34,200.
Where the US still dominates: capital and compute infrastructure. US corporate AI investment hit $344 billion in 2025. China's recorded figure was $12.4 billion. That is a 28-to-1 spending ratio. But spending does not equal performance — DeepSeek's V3 model proved that dramatically.
The takeaway is not that China "won." It is that the race is no longer a race. It is a tie at the frontier, with each country dominating different dimensions: the US in capital and chips, China in patents and robotics.
2. Coding Benchmarks Went From 60% to Near-Perfect in 12 Months
SWE-bench Verified is the industry standard for measuring whether AI can solve real-world software engineering tasks. In early 2025, the best models scored around 60%. By April 2026, top models approach 100%.
That is not incremental progress. That is the entire benchmark being effectively solved in a single year.
The implications cascade. SWE-bench tasks are derived from actual GitHub issues: real bugs in real codebases filed by real developers. A model scoring near-perfect on SWE-bench can fix most production bugs autonomously. It can implement feature requests from issue descriptions. It can read a failing test, trace the root cause, and write the patch.
Meanwhile, Humanity's Last Exam, designed as a ceiling-test of expert-level reasoning, went from 8.8% (OpenAI o1 in early 2025) to over 50% for the best models in April 2026. A benchmark explicitly designed to be too hard for AI was half-solved within a year of its creation.
But the report also reveals where AI still fails embarrassingly. The best model on ClockBench, a test of analog clock reading, scores just 50.6% (GPT-5.4). Claude Opus 4.6 manages 8.9%. AI can write production code better than most junior developers but cannot tell you what time a clock shows.
3. Junior Developer Employment Dropped 20% Since 2024
This is the number that should keep computer science departments awake at night. Employment among software developers aged 22 to 25 plummeted nearly 20% since 2024. Similar patterns appeared in customer service roles.
The report is careful to note correlation is not causation. But the pattern is unmistakable: entry-level positions decreased while mid-career and senior roles held steady or increased.
Here is the uncomfortable math. If AI models can now solve near-100% of SWE-bench tasks, the business case for hiring a junior developer to fix bugs and implement straightforward features weakens every quarter. Companies still need senior engineers to architect systems, review AI-generated code, and make judgment calls. But the on-ramp, the junior role that trains future seniors, is shrinking.
Global AI-related GitHub projects hit 5.58 million in 2025, a 23.7% year-over-year increase. More code is being written than ever. Just not by humans at the entry level.
4. AI Transparency Crashed While Adoption Soared
The Foundation Model Transparency Index tracks how much information AI companies disclose about their models: training data, parameter counts, safety testing, and compute costs. In 2025, the average score was 58 out of 100. In 2026, it dropped to 40.
That is a 31% decline in transparency in a single year.
Eighty of the 95 most notable models launched last year were released without training code. Google, Anthropic, and OpenAI have all stopped disclosing dataset sizes and training durations for their latest models. Over 90% of notable AI models now come from private companies.
At the same time, adoption is accelerating. Generative AI reached 53% global population adoption within three years, faster than the personal computer or the internet. 88% of organizations now use AI in some form.
The disconnect is stark. The tools are everywhere. Knowledge about how they work is disappearing.
5. Everyone Is Simultaneously Optimistic and Terrified
59% of people globally feel optimistic about AI benefits, up from 55% in 2024. At the same time, 52% report nervousness about AI products and services. Both numbers are rising.
This is not a contradiction. It is a perfectly rational response to a technology that makes your work 10x faster while threatening to eliminate your job.
The expert-public gap is widening. 73% of AI researchers and industry leaders are optimistic. Only 23% of the general public shares that level of confidence.
Trust in government regulation varies wildly. Singapore leads at 81%. The United States sits at the bottom with 31%.
Only 33% of Americans expect AI to improve their jobs, compared to 40% globally. The country that invented most of the frontier AI models trusts AI in the workplace less than the global average.
Full analysis with cross-links to related AI tools and open-source repos: news.skila.ai
Top comments (0)