How is the Value Index calculated?

Value = AvgIQ / (Input $/M + Output $/M). Both AvgIQ and Value are min–max normalized across the cohort before computing Unified.

US vs CHINA AI | About - Methodology & Scoring Explained

Q: What is the Unified Score?

The Unified Score (0–1000) emphasizes capability over cost efficiency: 10 × (0.9 × normalized AvgIQ + 0.1 × normalized Value). Avg IQ combines eighteen selected benchmarks using fixed allocations and a constant denominator. Missing results contribute zero. Each benchmark uses the highest verified score for the exact model across effort levels. Participation and benchmark coverage do not determine eligibility. Value is Avg IQ divided by total cost (Input $/M + Output $/M).

Q: What benchmarks are used to measure AI intelligence?

The selected benchmarks are Humanity’s Last Exam, SWE-bench Verified, GPQA, MMLU-Pro, Toolathlon, Agents’ Last Exam, DeepSWE, DeepSWE 1.1, LiveCodeBench v6, Terminal-Bench 2.1, BrowseComp, MMMU-Pro, CharXiv-R, ARC-AGI-2, OfficeQA Pro, IFBench, GPQA Diamond, FrontierMath — Tiers 1–3 v2. All eighteen are included regardless of participation. Missing results contribute zero; category aggregates are excluded to avoid double-counting.

Q: How is the National Score calculated?

National scores sum the Unified Scores for models from each nation in the combined leaderboard (Unified-desc).

What this site is

A leaderboard that ranks the world's best AI models by both intelligence and cost efficiency—then shows you which country is winning.

Use it to: Compare models side-by-side, find the best value for your use case, and see how the US-China AI race is unfolding in real time.

Why we built it

AI is a race—one that increasingly looks like a two-front contest between the United States and China. In a fast-moving landscape, debates tend to collapse into benchmark cherry-picking or price-only arguments. We wanted a simple, repeatable way to answer: “How strong is a model?” and “How much leverage do you get per dollar?”

This site turns that race into something you can track: who has the most capable models, who has the best economics, and who can place more models into the global Top 10.

🇺🇸 USA

Frontier capability, research leadership, and premium performance.

🇨🇳 China

Scaling, iteration speed, and cost-efficient deployment at volume.

The dashboard stays explicit about assumptions. If you disagree, you can still use the raw IQ and Value components directly.

Scoring model

Unified Score (0–1000)

Capability-focused (90%) with value efficiency (10%). Avg IQ combines eighteen selected benchmarks using fixed allocations, with missing results contributing zero.

Unified = 10 × (0.9 × norm(AvgIQ) + 0.1 × norm(Value))

Avg IQ: Combine the eighteen benchmark percentages using fixed allocations. The denominator always includes all eighteen, even when results are missing.
Best verified result: Use the highest score for each exact model and benchmark across all effort levels. Different benchmarks may use different effort levels. Preserve benchmark version, subset and protocol; exclude results that fall back to another model.
Coverage: No participation filters, participation multipliers or minimum number of reported benchmarks. Missing results receive zero contribution.
Value: AvgIQ / (Input $/M + Output $/M); models without usable pricing receive zero Value. Both Avg IQ and Value are cohort-normalized to 0–100 before blending into Unified.

National score (USA vs China)

National totals sum the Unified Scores for models from each nation in the combined leaderboard (sorted by Unified). This rewards both peak performance and depth of high-performing models.

Eligibility rule

Only models that appear in the global Top 10 contribute to a country’s national total.

Key assumptions

1) Intelligence is non-negotiable

Cheap output is not “value” if the model cannot solve hard problems.

2) Value contributes cost efficiency

A strong model becomes more powerful when it’s also affordable to deploy at scale.

3) The same eighteen benchmarks apply to every model

Avg IQ combines eighteen selected benchmarks using fixed allocations and a constant denominator. Missing results contribute zero. Each benchmark uses the highest verified score for the exact model across effort levels. Participation and benchmark coverage do not determine eligibility.

4) Pricing uses total cost

Value uses total cost (Input + Output) with linear normalization inside the Unified blend.

5) “All” views are for context

The headline national scores are based on the Top 10 filter, by design.

6) This is a snapshot, not a truth oracle

Benchmarks and pricing shift. Treat rankings as a periodic audit, not a permanent leaderboard.

Benchmarks included

All eighteen selected benchmarks are included, regardless of how many models report results:

GPQA MMLU-Pro Humanity’s Last Exam SWE-bench Verified Toolathlon Agents’ Last Exam DeepSWE DeepSWE 1.1 LiveCodeBench v6 Terminal-Bench 2.1 BrowseComp MMMU-Pro CharXiv-R ARC-AGI-2 OfficeQA Pro IFBench GPQA Diamond FrontierMath — Tiers 1–3 v2

Data notes

The model roster comes only from llm-stats. Missing benchmark results are researched from published, verifiable sources. Historical scores use the current scoring method with evidence available for each snapshot's date; higher results are never applied before their evidence availability date. Vendor pricing can change quickly; this is why the dashboard is presented as an audit at a point in time.

Not affiliated with any model provider. No guarantees are made about correctness, completeness, or fitness for any purpose.