An arena leaderboard over time — one line per model with a confidence band derived from its battle count, releases annotated with the move they produced, and a pairwise win-rate matrix that doubles as the card's control surface.
Preview in your theme
Loading preview…
This component is part of Pro
The install command, prompt and concept map are included with a Pro license — one purchase unlocks every Pro component, forever.
Standings and head-to-head win rates. Each row is a model, each column an opponent by rank, and the diagonal that model against the whole field. Arrow keys move the reading; the panel under the grid follows it.
standings · win rate vs →
1aurora-2-pro
2helix-4-thinking
3meridian-1.5-flash
4orion-mini-2026-03
5vega-70b-instruct
1aurora-2-pro1,462 ±3 · 41.8K battles
66%
52%
64%
70%
81%
2helix-4-thinking1,457 ±4 · 24.6K battles
48%
59%
63%
73%
–
3meridian-1.5-flash1,361 ±3 · 46.2K battles
36%
37%
52%
64%
69%
4orion-mini-2026-031,291 ±5 · 18.7K battles
30%
27%
36%
39%
61%
5vega-70b-instruct1,214 ±4 · 29.4K battles
19%
–
31%
39%
29%
aurora-2-pro vs the field
rank 1 of 5 · 1,462 ±3 · +58 since Jan 12 · 41,800 battles
52%won
21%tied
27%lost
14,700 battles of aurora-2-pro's 41,800 against 4 opponents on this card. aurora-2-pro takes 65.5% of the 11,634 decided ones; the current ratings predict 65.8%, a residual of -0.2 points.
The band is 1.96 standard errors of the rating, and the standard error comes from the battle count alone: a Bradley–Terry fit takes n/4 units of information from n comparisons at even odds, so SE = 2 · 400/ln10 ÷ √n — ±5 at 18,700 battles, ±3 at 46,200. 3 of 4 neighbouring gaps on the ladder clear both models' bands; aurora-2-pro and helix-4-thinking are statistically tied. 9 of 10 pairs have battles recorded. The matrix and the ladder disagree most on meridian-1.5-flash over orion-mini-2026-03: 64.4% observed against 59.9% implied by the ratings, +4.5 points. Click a standings row or a cell, or Tab into the grid and use the arrow keys; the diagonal reads a model against the whole field.
Arena ratings for 5 models, Jan 12, 2026 to Jun 15, 2026. aurora-2-pro leads at 1,462 ±3 on 41,800 battles. 3 of 4 neighbouring gaps on the ladder clear both models' bands. Exact values follow in the data table. Up to 12 snapshot days, evenly spaced by rank; each cell is the rating with its 1.96-sigma band and the battles behind it.
Date
aurora-2-pro
helix-4-thinking
meridian-1.5-flash
orion-mini-2026-03
vega-70b-instruct
Jan 12, 2026
1,404 ±9, 6,200 battles
not rated on this day
1,330 ±7, 9,400 battles
not rated on this day
1,210 ±9, 5,600 battles
Jan 26, 2026
1,412 ±7, 9,800 battles
not rated on this day
1,333 ±6, 13,100 battles
not rated on this day
1,208 ±7, 8,300 battles
Feb 9, 2026
1,424 ±6, 13,400 battles
not rated on this day
1,335 ±5, 17,000 battles
not rated on this day
1,213 ±6, 11,000 battles
Feb 23, 2026
1,447 ±5, 17,600 battles
not rated on this day
1,334 ±5, 20,800 battles
not rated on this day
1,211 ±6, 13,600 battles
Mar 9, 2026
1,449 ±5, 21,500 battles
1,391 ±16, 1,900 battles
1,336 ±4, 24,300 battles
1,288 ±15, 2,100 battles
1,209 ±5, 16,100 battles
Mar 23, 2026
1,452 ±4, 25,100 battles
1,416 ±10, 4,300 battles
1,338 ±4, 27,900 battles
1,284 ±10, 4,600 battles
1,212 ±5, 18,500 battles
Apr 6, 2026
1,450 ±4, 28,400 battles
1,444 ±8, 8,200 battles
1,337 ±4, 31,200 battles
1,290 ±8, 7,500 battles
1,215 ±5, 20,800 battles
Apr 20, 2026
1,454 ±4, 31,600 battles
1,449 ±6, 12,100 battles
1,339 ±4, 34,600 battles
1,287 ±7, 10,200 battles
1,213 ±4, 23,000 battles
May 4, 2026
1,457 ±4, 34,500 battles
1,452 ±5, 15,600 battles
1,340 ±4, 37,800 battles
1,292 ±6, 12,600 battles
1,211 ±4, 25,100 battles
May 18, 2026
1,459 ±4, 37,200 battles
1,454 ±5, 18,900 battles
1,358 ±3, 40,900 battles
1,289 ±6, 15,000 battles
1,214 ±4, 27,100 battles
Jun 1, 2026
1,460 ±3, 39,700 battles
1,456 ±5, 21,900 battles
1,360 ±3, 43,700 battles
1,290 ±5, 17,000 battles
1,212 ±4, 28,900 battles
Jun 15, 2026
1,462 ±3, 41,800 battles
1,457 ±4, 24,600 battles
1,361 ±3, 46,200 battles
1,291 ±5, 18,700 battles
1,214 ±4, 29,400 battles
empty
Arena ratings · text arena, 2026 H1
Arena ratings over time, banded by the battles behind each estimate
No rated snapshots
A ratings card needs at least one model carrying a dated snapshot of rating and battles. The band, the standings and every separation test are derived from that pair — a leaderboard without battle counts can be ranked but cannot be tested.
loading
Arena ratings · text arena, 2026 H1
Arena ratings over time, banded by the battles behind each estimate
Loading arena ratings
error
Arena ratings · text arena, 2026 H1
Arena ratings over time, banded by the battles behind each estimate
Zyeon UI