Frontier model leaderboard.

Rows are sorted by solve rate (Reward %): the share of tasks a model solves (mean rubric ≥ 3.0/5 AND checklist coverage ≥ 60%, experts averaged per task). Small grey numbers are 95% bootstrap CIs (1,000 resamples). Judgments are not independent (two correlated judgments per task), so every CI is clustered over the 598 tasks. Overlapping intervals are a conservative check.

#	Model	Reward %	Overall	Dom.	Strat.	Act.	Exec.	Local Fid.	Cov.	Avg Rank	Win %	Top-3 %
01	Fable 5	58%54–62	83%82–84	84%	86%	83%	85%	77%	63%	2.133	50%47–53	82%
02	GPT-5.5	52%48–56	78%77–79	82%	81%	81%	75%	70%	59%	2.849	22%20–25	69%
03	Claude Opus 4.8	36%33–40	76%75–77	76%	80%	74%	82%	66%	53%	3.242	9%8–11	61%
04	Gemini 3.1 Pro	27%23–30	72%71–73	71%	73%	74%	79%	62%	48%	3.778	7%6–9	38%
05	GLM-5.2	23%20–26	68%67–69	66%	69%	72%	77%	59%	44%	4.156	6%5–7	30%
06	Mistral Large 3	19%16–22	61%60–62	58%	60%	68%	69%	52%	40%	4.842	6%4–7	20%

Blind preference distribution

Where each model lands across all blind rankings, from #1 (best) to #6.

Fable 5

#1 in 50%

GPT-5.5

#1 in 22%

Claude Opus 4.8

#1 in 9%

Gemini 3.1 Pro

#1 in 7%

GLM-5.2

#1 in 6%

Mistral Large 3

#1 in 6%

#1 rank#2 rank#3 rank#4 rank#5 rank#6 rank

Head-to-head

Each cell is the row model's win rate over the column model, computed only on the blind rankings where an expert placed both. Blue = the row wins the majority of those matchups; red = it loses. Beats counts how many opponents each model beats outright.

The ranking is fully transitive: the "Beats" column runs 5, 4, 3, 2, 1, 0, so each model beats every model below it and loses to every model above, with no cycles.

Model	Fable 5	GPT-5.5	Opus 4.8	Gemini 3.1 Pro	GLM-5.2	Mistral Large 3	Beats
Fable 5	—	66%	76%	80%	82%	82%	5/5
GPT-5.5	34%	—	56%	68%	74%	83%	4/5
Claude Opus 4.8	24%	44%	—	62%	69%	77%	3/5
Gemini 3.1 Pro	20%	32%	38%	—	58%	74%	2/5
GLM-5.2	18%	26%	31%	42%	—	68%	1/5
Mistral Large 3	18%	17%	23%	26%	32%	—	0/5