Score stability 1.0

Models: 18
Smallest CI width: 1.86
Largest CI width: 10.81
Most stable: Claude Opus 5 (high)

Two-dimensional view

Stability.

Each dot is one model. Lower means a better average question score. Further left means a narrower CI, so the model's results varied less across repeated rounds on these subjects.

Lower-left is betterNumbers match the table rank. Select a point to open the model's results.

1. Claude Opus 5 (high)12.34 questions · CI width 1.86
2. GPT-5.6 Sol (high)14.40 questions · CI width 3.11
3. Gemini 3.8 Flash (high)13.14 questions · CI width 3.25
4. GPT-5 Nano (medium)15.49 questions · CI width 3.29
5. Claude Fable 5 (high)12.06 questions · CI width 3.85
6. Grok 4.6 (high)14.29 questions · CI width 3.85
7. Claude Sonnet 5 (high)14.74 questions · CI width 3.93
8. Gemini 3.6 Flash (high)16.89 questions · CI width 5.61
9. Ox Alpha (high)17.60 questions · CI width 5.99
10. GPT-5.6 Luna (high)17.74 questions · CI width 6.01
11. GPT-6 Astra (high)15.09 questions · CI width 6.42
12. Gemini 3.7 Flash (high)14.03 questions · CI width 6.49
13. Claude Fable 5.1 (high)12.11 questions · CI width 6.60
14. Grok 4.5 (high)15.17 questions · CI width 6.94
15. Kimi K3 (high)12.74 questions · CI width 7.18
16. gpt-oss-120B (high)13.63 questions · CI width 7.67
17. Mistral Medium 3.5 (high)20.06 questions · CI width 7.90
18. Llama 4 Maverick (non-thinking)32.23 questions · CI width 10.81

Stability ranking
Stability rank					95% CI
Rank 1	Claude Opus 5 (high)anthropic	1.86	3	12.34	11.41–13.27	100%
Rank 2	GPT-5.6 Sol (high)openai	3.11	9	14.40	12.84–15.96	100%
Rank 3	Gemini 3.8 Flash (high)google-ai-studio	3.25	5	13.14	11.52–14.77	100%
Rank 4	GPT-5 Nano (medium)openai	3.29	13	15.49	13.84–17.13	94%
Rank 5	Claude Fable 5 (high)anthropic	3.85	1	12.06	10.13–13.98	100%
Rank 6	Grok 4.6 (high)xai	3.85	8	14.29	12.36–16.21	100%
Rank 7	Claude Sonnet 5 (high)anthropic	3.93	10	14.74	12.78–16.71	100%
Rank 8	Gemini 3.6 Flash (high)google-vertex	5.61	14	16.89	14.08–19.69	97%
Rank 9	Ox Alpha (high)stealth	5.99	15	17.60	14.60–20.60	91%
Rank 10	GPT-5.6 Luna (high)openai	6.01	16	17.74	14.74–20.75	91%
Rank 11	GPT-6 Astra (high)openai	6.42	11	15.09	11.88–18.30	94%
Rank 12	Gemini 3.7 Flash (high)google-ai-studio	6.49	7	14.03	10.78–17.28	97%
Rank 13	Claude Fable 5.1 (high)anthropic	6.60	2	12.11	8.81–15.41	97%
Rank 14	Grok 4.5 (high)xai	6.94	12	15.17	11.70–18.64	94%
Rank 15	Kimi K3 (high)moonshotai	7.18	4	12.74	9.16–16.33	94%
Rank 16	gpt-oss-120B (high)cerebras	7.67	6	13.63	9.79–17.46	94%
Rank 17	Mistral Medium 3.5 (high)mistral	7.90	17	20.06	16.11–24.01	89%
Rank 18	Llama 4 Maverick (non-thinking)parasail	10.81	18	32.23	26.82–37.63	49%

Claude Opus 5 (high)anthropic1.86CI width

Question score: 12.34
95% CI: 11.41–13.27
Solved: 100%

GPT-5.6 Sol (high)openai3.11CI width

Question score: 14.40
95% CI: 12.84–15.96
Solved: 100%

Gemini 3.8 Flash (high)google-ai-studio3.25CI width

Question score: 13.14
95% CI: 11.52–14.77
Solved: 100%

GPT-5 Nano (medium)openai3.29CI width

Question score: 15.49
95% CI: 13.84–17.13
Solved: 94%

Claude Fable 5 (high)anthropic3.85CI width

Question score: 12.06
95% CI: 10.13–13.98
Solved: 100%

Grok 4.6 (high)xai3.85CI width

Question score: 14.29
95% CI: 12.36–16.21
Solved: 100%

Claude Sonnet 5 (high)anthropic3.93CI width

Question score: 14.74
95% CI: 12.78–16.71
Solved: 100%

Gemini 3.6 Flash (high)google-vertex5.61CI width

Question score: 16.89
95% CI: 14.08–19.69
Solved: 97%

Ox Alpha (high)stealth5.99CI width

Question score: 17.60
95% CI: 14.60–20.60
Solved: 91%

GPT-5.6 Luna (high)openai6.01CI width

Question score: 17.74
95% CI: 14.74–20.75
Solved: 91%

GPT-6 Astra (high)openai6.42CI width

Question score: 15.09
95% CI: 11.88–18.30
Solved: 94%

Gemini 3.7 Flash (high)google-ai-studio6.49CI width

Question score: 14.03
95% CI: 10.78–17.28
Solved: 97%

Claude Fable 5.1 (high)anthropic6.60CI width

Question score: 12.11
95% CI: 8.81–15.41
Solved: 97%

Grok 4.5 (high)xai6.94CI width

Question score: 15.17
95% CI: 11.70–18.64
Solved: 94%

Kimi K3 (high)moonshotai7.18CI width

Question score: 12.74
95% CI: 9.16–16.33
Solved: 94%

gpt-oss-120B (high)cerebras7.67CI width

Question score: 13.63
95% CI: 9.79–17.46
Solved: 94%

Mistral Medium 3.5 (high)mistral7.90CI width

Question score: 20.06
95% CI: 16.11–24.01
Solved: 89%

Llama 4 Maverick (non-thinking)parasail10.81CI width

Question score: 32.23
95% CI: 26.82–37.63
Solved: 49%

How it is calculated

CI width.

A smaller CI width means more consistent results across the repeated rounds.

FormulaCI width = upper 95% CI bound − lower 95% CI bound

Calculation detailsSteps, example, and limits

Calculate the model's 95% CI from the spread between repeated rounds of the same subject. Differences between subjects do not widen it.
Subtract the lower end from the upper end.
Sort the widths from smallest to largest.

Example, Claude Opus 5 (high): 13.27 − 11.41 = 1.86 questions.

Stable does not mean good. A model that scores poorly but consistently can rank well here. A model with a strong average and a wide CI ranks lower.

This is an approximate measure for these subjects only. It does not predict single rounds and is not a significance test between two models.