How the scoreboard's 0-100 score works
Every lab on the board gets a single score from 0 to 100. It is a weighted average of up to five components, each already scaled to 0-100. If a lab is missing a component, that component is dropped and the remaining weights are rescaled, so every lab still gets a 0-100 score even when the inputs aren't equally complete. The site shows "coverage" (e.g. "5/5" or "3/5") next to every score so you can see how much data it's built from.
The five components and their weights
| Component | Weight | Formula |
|---|---|---|
| Arena | 35 | 100 x (score - (leader - 100)) / 100, clamped 0-100 |
| Artificial Analysis | 15 | 100 x index / best index |
| Revenue | 20 | 100 x log10(revenue / $100M) / log10(top revenue / $100M) |
| Valuation | 10 | 100 x log10(valuation / $1B) / log10(top valuation / $1B) |
| Polymarket odds | 20 | 100 x price / highest price |
The weights sum to 100. Arena, the crowd-voted text leaderboard, carries the single biggest weight. Revenue and valuation use a log scale rather than a linear one, because the gap between, say, a $1B lab and a $10B lab matters more than the same $9B gap between a $100B lab and a $109B lab.
The overall formula: score = sum(weight_i x component_i) / sum(weight_i), summed only over the components a lab actually has.
A full-coverage example: Anthropic
As of 19 Oct 2026, Anthropic has all five components:
- Arena: score 1508, model Claude Opus 5.5 (High), as of 15 Oct 2026. The leader (Google DeepMind) scores 1524. Component = 1508 - (1524 - 100) = 84.
- Artificial Analysis: index 57.6, the highest of any tracked lab, so component = 100.
- Revenue: $67.2B annualized, the highest tracked figure, so component = 100.
- Valuation: $965B post-money (28 May 2026), the highest tracked figure, so component = 100.
- Polymarket odds: 49.5% Yes price on "best AI model end of 2026," the highest among labs in that market, so component = 100.
Weighted sum: (35 x 84) + (15 x 100) + (20 x 100) + (10 x 100) + (20 x 100) = 2940 + 1500 + 2000 + 1000 + 2000 = 9440. Divide by the full weight of 100: 9440 / 100 = 94.4, matching Anthropic's published score exactly.
A rescaled example: Google DeepMind
Google DeepMind is a division of a public company, Alphabet, which does not disclose lab-level revenue or valuation. So Google is scored on only 3 of 5 components: Arena, Artificial Analysis and odds.
- Arena: score 1524, which is the leader, so component = 1524 - (1524 - 100) = 100.
- Artificial Analysis: index 52.56, versus the top index of 57.62 (Anthropic). Component = 100 x 52.56 / 57.62 = 91.2.
- Polymarket odds: 39.5%, versus the top price of 49.5% (Anthropic). Component = 100 x 39.5 / 49.5 = 79.8.
The available weight is only 35 + 15 + 20 = 70 (revenue's 20 and valuation's 10 are dropped). Weighted sum: (35 x 100) + (15 x 91.2) + (20 x 79.8) = 3500 + 1368 + 1596 = 6464. Divide by 70: 6464 / 70 = 92.3, again matching the published score.
Because the remaining weights are rescaled rather than left as zero, a lab with partial coverage isn't automatically punished for missing components, it's judged purely on what it does have. That's a deliberate design choice, and also a known bias: it tends to help labs that sit inside bigger public companies, since they never have to clear the revenue and valuation bars at all. The methodology page calls this out directly.
Sources
- Arena text leaderboard (opens in new tab) (as of 15 Oct 2026)
- Artificial Analysis leaderboards (opens in new tab) (fetched 19 Oct 2026)
- Which company has best AI model end of 2026? (opens in new tab) (Polymarket, fetched 19 Oct 2026)
- Anthropic: Series H announcement (opens in new tab) (28 May 2026)
- Anthropic's third-quarter revenue hits $16.8 billion (opens in new tab) (CNBC, 15 Oct 2026)