Stanford AI Index 2026: China Has Nearly Closed the US Performance Gap

By Hector Herrera | April 15, 2026 | Science

The 2026 Stanford AI Index, released April 13, documents two trends that should concern anyone tracking the AI industry: China's top models have nearly matched US leaders in raw performance, and the major AI labs are becoming dramatically less transparent about how their models work. Neither trend is moving in a good direction.

These are not predictions. They are measurements.

Background: What the Stanford AI Index Is

The Stanford AI Index is an annual report produced by Stanford's Human-Centered AI Institute (HAI). Now in its eighth year, it tracks AI performance benchmarks, investment, research output, policy developments, and increasingly, AI safety and transparency metrics. It is the most comprehensive annual snapshot of where the field stands, drawing on data from academic institutions, industry, and government sources.

The Performance Gap Is Closing Fast

In early 2024, the best US model outperformed China's best by 9.26 percentage points on standardized benchmarks. By early 2025 — one year later — that gap had narrowed to 1.70 percentage points.

That is not incremental progress. That is a near-total closure of the performance lead in 12 months.

The 2026 report notes the gap has continued to narrow since that measurement. The specific benchmarks involved are composite measures across reasoning, language, coding, and multimodal tasks — the same categories that define competitive advantage for commercial AI products.

What drove the convergence:

• Chinese labs including DeepSeek, Baidu, and Zhipu AI significantly increased compute investment and training scale
• Open-weight model releases (including Meta's Llama series) provided starting points that reduced the research gap
• Adversarial distillation techniques — now the subject of a US lab coordination effort — may have accelerated capability acquisition
• US export controls on advanced chips created pressure for Chinese labs to optimize efficiency, producing models that perform competitively on less compute

The Transparency Crisis

The second major finding is less covered but arguably more important for the long term: AI transparency is collapsing.

The Foundation Model Transparency Index — a measure of how openly AI labs disclose information about their models — fell from an average score of 58 to 40 points between 2024 and 2025. That is a 31% drop.

What transparency includes:

• Training data — what datasets were used, what was excluded, and how data was filtered
• Parameter counts — how large the model is (directly related to its capabilities and compute requirements)
• Model architecture — the structural design choices that define how the model processes information
• Evaluation methodology — how the lab measured its own model's performance

Labs are increasingly treating all of this as proprietary information. The competitive pressure is understandable — if you disclose your training data, competitors can replicate your approach. But the consequences extend beyond business competition:

1. Independent safety evaluation becomes harder. Researchers who want to test whether a model has dangerous capabilities need to understand how it was built.
2. Benchmark credibility weakens. If labs control both the model and the evaluation methodology, self-reported performance numbers are harder to trust.
3. Regulatory oversight is undermined. Policymakers cannot write effective rules for systems they cannot see inside.

US Policy Implications

The Stanford report arrives at a politically charged moment. The 1.70% performance gap gives ammunition to hawks who argue export controls and investment restrictions are working but insufficient. It gives ammunition to skeptics who argue the controls are failing.

The more useful framing: the gap has narrowed, the tools China used to close it are partially known, and the US response so far has been reactive. The coalition between OpenAI, Anthropic, and Google to block adversarial distillation is a defensive measure. Export controls on chips are a supply-side constraint. Neither directly addresses the underlying dynamic: China is running a well-resourced, sustained AI development program, and the performance convergence reflects that investment.

What the data does not tell you:

• Whether the remaining 1.70% gap is meaningful for real-world applications (it may be, in specific domains)
• Whether Chinese models have equivalent performance on safety and alignment — a dimension the transparency collapse makes harder to assess
• Whether the gap will widen again as US labs deploy new frontier models including Anthropic's Claude Mythos and Google's next Gemini release

What to Watch

The 2026 Stanford AI Index will be cited extensively in congressional testimony through the spring and summer. Watch for it to appear in Senate Armed Services and Intelligence Committee hearings on AI and national security. The transparency findings are likely to inform proposed reporting requirements for large AI training runs — a provision that has appeared in multiple legislative drafts and keeps getting dropped. This report strengthens the case for including it.

Hector Herrera is the founder of Hex AI Systems and editor of NexChron.