Methodology
How the score works, where the data comes from, and what we don't know.
The score (0–100)
Each lab gets up to five component scores, each 0–100. The overall score is their weighted average. If a component has no reliable public data, it is dropped and the remaining weights are rescaled. We never guess. Each score shows its coverage (e.g. "3/5").
| Component | Weight | Formula |
|---|---|---|
| arena | 35 | Arena text score: 100 × (score − (leader − 100)) / 100, clamped to 0–100. The leader gets 100; 100 Elo points behind gets 0. |
| aa | 15 | Artificial Analysis Intelligence Index: 100 × index / best index. |
| revenue | 20 | Revenue run-rate on a log scale: 100 × log10(rev / $100M) / log10(top rev / $100M). $100M = 0. |
| valuation | 10 | Latest closed valuation on a log scale: 100 × log10(val / $1B) / log10(top val / $1B). $1B = 0. |
| odds | 20 | Polymarket "best AI model end of 2026" Yes price: 100 × price / highest price. |
score = Σ (weight_i × component_i) / Σ weight_i (over available components only)
Known bias: rescaling helps labs inside public giants (Google, Meta), which are scored only on models and odds. The coverage label is there so you can see it. Compute (GW, capex) is shown on lab pages but not scored, because the figures aren't comparable across labs.
Sources & refresh
- Polymarket: public Gamma API, refreshed automatically. The scored market is "Which company has best AI model end of 2026?", which itself resolves on the Arena text leaderboard.
- Artificial Analysis: public leaderboard page, refreshed automatically. "Position" is the rank among non-deprecated models that have an Intelligence Index on that page, not an official AA rank.
- Arena (formerly LMArena): text leaderboard, Overall. The site blocks automated access, so we use a manually copied snapshot (currently 15 Oct 2026) and flag it when it's older than 8 days.
- Revenue / valuation / funding: no API exists. These are updated by hand from filings, company posts and major outlets, with a news-alert helper to catch changes. Each figure shows its basis, as-of date and link.
Claimed vs reported
When a company or its investors push a higher number than what is later reported (e.g. a "grossed-up" annualized figure), we show both and score the reported one. Run-rates use each company's own accounting. Anthropic books some cloud-partner revenue gross and OpenAI books it net, so the figures are not perfectly comparable. Where we annualize a quarter ourselves, it says "q × 4".
Weekly changes
Each weekly run saves a snapshot to /data/history/. Arrows compare with the snapshot from about 7 days earlier. The first comparison week (10 Oct 2026) is backfilled: its odds are real historical Polymarket prices, and its financials are limited to facts dated on or before that day. Arena and AA values were held at current levels, because no history was captured.
Independence & disclaimer
Sponsors and advertisers never affect scores or rankings. This site is not affiliated with any AI lab, Arena, Artificial Analysis or Polymarket. Nothing here is investment, trading or betting advice. Data can be wrong or stale. Every figure links to its source, so check it. Prediction-market prices are probabilities implied by traders, not facts. Spot an error? It will be fixed and noted in the lab's timeline.