Frontier model leaderboard

A blind expert evaluation of six frontier models over 200 executive-reasoning items, ranked by Reward, the share of prompts a model solves. The rubric dimensions, win rate, and rank distributions that follow are the diagnostic detail behind that ranking.

Reward: the share of prompts solved

A prompt counts as solved only when the experts give the answer a mean rubric score of at least 3.0/5 and it covers at least 60% of the checklist, so Reward credits substance over presentation. Fable 5 leads at 52%, and the field falls to 12%.

Fable 5

52%

GPT-5.5

43%

Claude Opus 4.8

30%

Gemini 3.1 Pro

22%

GLM-5.2

20%

Mistral Large

12%

Full leaderboard i Reward % share of prompts solved (mean rubric ≥ 3/5 AND coverage ≥ 60%), the headline score.
Overall mean of the five rubric dimensions, as a % of the 5-point max.
Dom / Strat / Act / Exec / Local the five rubric dimensions, each a % of the 5-point max.
Cov. mean checklist coverage.
Avg rank mean blind rank (1 = best).
Win % share of blind judgments ranking the model first.
Top-3 % share ranking it top three.
Small grey numbers are 95% bootstrap confidence intervals.

Rows are sorted by Reward %: the share of prompts a model solves (mean rubric ≥ 3.0/5 AND checklist coverage ≥ 60%, experts averaged per item). Small grey numbers are 95% bootstrap CIs (1,000 resamples). The 400 judgments per model are not independent (two correlated judgments per item), so the effective sample is 200 items, not 400, and CIs are clustered over items accordingly. Overlapping CIs mean the separation is not significant.

#	Model	Reward %	Overall	Dom.	Strat.	Act.	Exec.	Local Fid.	Cov.	Avg rank	Win %	Top-3 %
1	Fable 5	52% 44–57	82% 80–83	84%	76%	85%	83%	82%	59%	2.103	50% 45–55	83%
2	GPT-5.5	43% 36–50	77% 75–78	82%	68%	80%	73%	81%	55%	2.877	22% 18–27	68%
3	Claude Opus 4.8	30% 24–37	75% 73–76	76%	63%	79%	80%	74%	49%	3.223	8% 5–10	63%
4	Gemini 3.1 Pro	22% 16–28	71% 69–73	70%	62%	72%	78%	74%	45%	3.77	8% 6–11	37%
5	GLM-5.2	20% 14–26	67% 66–69	66%	56%	67%	75%	72%	41%	4.16	6% 4–9	30%
6	Mistral Large	12% 8–17	59% 58–61	57%	49%	57%	65%	69%	35%	4.867	6% 4–8	19%

Blind preference distribution

Where each model lands across all 400 blind rankings, from #1 (best) to #6.

Fable 5#1 in 50%

GPT-5.5#1 in 22%

Claude Opus 4.8#1 in 8%

Gemini 3.1 Pro#1 in 8%

GLM-5.2#1 in 6%

Mistral Large#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. Violet = the row wins the majority of those matchups; red = it loses. Beats counts how many opponents each model beats outright.

The ranking is perfectly transitive: the “Beats” column runs 5, 4, …, 0 with no ties, so every model beats exactly the ones below it and loses to those above, with no rock-paper-scissors cycles. That is a strong internal-consistency signal for the ordering, independent of the rubric means.

Model	Fable	GPT-5.5	Claude	Gemini	GLM-5.2	Mistral	Beats
Fable 5	—	66%	76%	81%	82%	84%	5/5
GPT-5.5	34%	—	55%	67%	74%	84%	4/5
Claude Opus 4.8	24%	46%	—	62%	69%	77%	3/5
Gemini 3.1 Pro	19%	33%	38%	—	60%	74%	2/5
GLM-5.2	18%	26%	31%	40%	—	69%	1/5
Mistral Large	16%	16%	23%	26%	31%	—	0/5

Mistral Large

69%

Rubric scores by domain

Mean rubric score (% of the 5-point max) by executive domain. Darker = stronger. Per-domain n is small (esp. Finance): read these as directional, not precise.

Model	Business	Finance	Marketing	Product & Tech
Fable 5	79%	91%	82%	79%
GPT-5.5	75%	76%	80%	77%
Claude Opus 4.8	72%	80%	75%	74%
Gemini 3.1 Pro	68%	78%	71%	69%
GLM-5.2	67%	68%	69%	66%
Mistral Large	63%	54%	62%	58%

Explore one graded item end to end in the sample item, or see how the set is built on the methodology page.