Methodology

Overview

The AI Race scores organizations across six key pillars, each capturing a different dimension of AI leadership. Scores are grounded in real benchmark data — we track 24 public benchmarks across 7 categories, automatically ingested daily from authoritative sources. This is supplemented by product launches, partnerships, research papers, and market adoption metrics. Each pillar is scored from 0 to 100.

Six Pillars of Assessment

Benchmark Data

We track 24 public benchmarks across 7 categories. Each benchmark is normalized to a 0–100 scale and weighted within its category. ELO-based scores use a linear mapping from 800–1400 ELO to 0–100. Percentage-based scores map directly.

Modality Scoring

In addition to overall scores, each organization is assessed across four modalities: Text, Image, Video, and Multimodal. These are computed from the benchmark categories above and contribute to the category-specific leaderboards.

Confidence Levels

Each score is assigned a confidence level based on data availability:

Score Computation

The overall score is a weighted average of the six pillar scores:

The Capability pillar is informed by benchmark data when available. Category scores are computed as weighted averages of normalized benchmark scores within each category, re-normalized to account for missing data points.

Rank movement is computed by comparing each organization's rank in the latest snapshot against the previous snapshot. A positive delta means the org moved up.

Data Sources

Benchmark data is automatically ingested from these primary sources:

• •LM Arena / Chatbot Arena — ELO ratings for text, image, and video models via crowdsourced blind comparisons
• •Artificial Analysis — standardized benchmark scores (MMLU-Pro, GPQA, SimpleQA, MATH-500, HumanEval+, IFEval)
• •HuggingFace Datasets — open datasets for arena scores and evaluation results

Scores are also informed by:

• •Official blog posts, technical reports, and research papers
• •Press releases, funding announcements, and SEC filings
• •App store rankings, user metrics, and adoption data
• •Hugging Face model statistics, GitHub activity, and developer adoption

Update Cadence

Benchmark data is automatically refreshed daily at 6:00 AM UTC via our ingestion pipeline, which pulls the latest scores from LM Arena, Artificial Analysis, and HuggingFace. Editorial scores (velocity, adoption, compute, ecosystem, trust) are updated approximately every 2–4 weeks, or sooner when major events occur (significant model releases, large funding rounds, benchmark-shifting results). Each update is documented in the changelog.