Why Arena and Artificial Analysis disagree
The scoreboard pulls two separate model-quality signals: Arena's text leaderboard and the Artificial Analysis Intelligence Index. They frequently disagree about who is "best," and the disagreement is informative, not noise.
The two measurements are different things
Arena's score is built from head-to-head human preference votes: people see two anonymous model outputs side by side and pick the one they like better, and the scores behave like a ranking (Elo-style) built from those pairwise results. The scoreboard's Arena snapshot is manual and dated, as of 15 Oct 2026, and gets flagged on the site when it's older than 8 days.
The Artificial Analysis Intelligence Index aggregates a battery of benchmark evaluations into a single number, refreshed automatically (fetched 19 Oct 2026 for this comparison, covering 269 models). Its "position" is explicitly not an official Artificial Analysis rank: the methodology defines it as "the rank among non-deprecated models with an Intelligence Index on that page."
One measures human preference; the other is a benchmark composite. A model can win one and lose the other.
Where the two leaderboards split, as of 19 Oct 2026
| Lab | Arena rank | Arena score | Arena model | AA position | AA index | AA model |
|---|---|---|---|---|---|---|
| Google DeepMind | 1 | 1524 | Gemini 4 Argon (High) | 9 | 52.6 | Gemini 4 Argon (High) |
| Meta | 2 | 1509 | Muse Wonder (Max) | 4 | 55.8 | Muse Wonder Full (Max) |
| Anthropic | 3 | 1508 | Claude Opus 5.5 (High) | 1 | 57.6 | Claude Opus 5.5 (Max, Default Fallback) |
| OpenAI | 21 | 1485 | GPT-5.6 Sol (xHigh) | 8 | 52.7 | GPT-6 Astra (Max) |
Google DeepMind is the starkest case: #1 on Arena, but 9th on Artificial Analysis. Anthropic is the mirror image: 1st on Artificial Analysis, but 3rd on Arena. Meta sits 2nd on Arena and 4th on AA, closer to consistent, but still not matching.
OpenAI's row shows another wrinkle: its best-placed model on Arena, GPT-5.6 Sol, is not the release Artificial Analysis rates highest, GPT-6 Astra. The board takes each lab's best row on each leaderboard, so the two can credit different models from the same lab, which is itself a reason the two rankings can diverge.
How this plays into the overall score
Arena carries the largest single weight on the scoreboard, 35 of 100, versus 15 for Artificial Analysis. So a lab's Arena position has more influence on its overall score than its AA position does. That is part of why Google DeepMind's Arena #1 keeps it competitive overall (component score 100, since it is the Arena leader) despite ranking 9th on AA (component score 91.2). It is also part of why Anthropic's AA #1 (component score 100) isn't by itself enough to overtake Google on the model-quality side of the formula: Anthropic's Arena component is 84, lower than Google's 100.
Anthropic still outscores Google overall, 94.4 to 92.3, as of 19 Oct 2026. Two things do it: Anthropic leads the Polymarket odds component (100 vs Google's 79.8), and it scores 100 on revenue and valuation, which Google has no standalone figures for (those weights are dropped for Google, not counted as zero).
Sources
- Arena text leaderboard (opens in new tab) (as of 15 Oct 2026)
- Artificial Analysis leaderboards (opens in new tab) (fetched 19 Oct 2026)