secops-es-benchmark — leaderboard

SecOps investigation agents scored on real, labeled Elasticsearch telemetry. Ranked by overall score (mean of objective & tasks); higher is better (0–100).

Objective = 54 auto-graded questions (deterministic). Tasks = 5 investigations, LLM judge: claude-opus-5 (report-only re-judge). Same read-only tool surface for every model. Agents ran with extended thinking OFF (Claude) / provider default (OpenAI-compatible endpoints) — this can understate reasoning-heavy models. The one (thinking) row is the same model re-run with extended thinking ON, for comparison. Single run per model. Generated 2026-08-03.

Objective % Tasks %

OverallObjectiveTasks

1claude-opus-4-8 (thinking)73.3

78.3

68.4

2claude-opus-4-875.5

73.1

77.9

3claude-sonnet-4-575.2

87.8

62.6

4qwen3.7-plus67.8

72.3

63.4

5openai/gpt-5.6-sol65.5

54.5

76.5

6glm-5.259.3

72.6

46.1

7qwen3.6-27b54.9

73.4

36.3

8google/gemma-4-31b-it46.5

61.1

9gemma-4-26b-a4b-it-heretic-guff46.3

60.8

31.8

10claude-haiku-4-545.2

63.7

26.6

11minimax-m2.7-heretic (MLX 8bit)42.8

54.3

31.3

12minimax-m2.5 (MLX 8bit)41.1

51.4

30.9

Objective by difficulty (%)

	easy	medium	hard
claude-opus-4-8 (thinking)	90	82.7	63.3
claude-opus-4-8	70	76.2	69.5
claude-sonnet-4-5	100	91.4	73.8
qwen3.7-plus	80	70.9	69.8
openai/gpt-5.6-sol	70	56.3	41.6
glm-5.2	90	75	57.6
qwen3.6-27b	70	73.9	74.6
google/gemma-4-31b-it	50	69.6	53.1
gemma-4-26b-a4b-it-heretic-guff	60	61	61.2
claude-haiku-4-5	50	66.9	66.8
minimax-m2.7-heretic (MLX 8bit)	40	60.7	51.9
minimax-m2.5 (MLX 8bit)	30	54.3	59.7

Objective by question type (%)

	boolean	extraction	labeling	mcq	ordering	set
claude-opus-4-8 (thinking)	75	90	31.1	100	90	49.2
claude-opus-4-8	50	80	41.8	100	91.7	56.6
claude-sonnet-4-5	100	90	70.1	100	85	64.9
qwen3.7-plus	75	70	60	100	68.3	45.7
openai/gpt-5.6-sol	50	60	35.2	63.6	85	36.9
glm-5.2	62.5	80	26.2	100	66.7	57.2
qwen3.6-27b	62.5	70	69.7	100	78.3	57.1
google/gemma-4-31b-it	62.5	55	37.6	90.9	41.7	53.5
gemma-4-26b-a4b-it-heretic-guff	62.5	50	54.9	100	75	32.7
claude-haiku-4-5	75	50	47.2	100	63.3	47.4
minimax-m2.7-heretic (MLX 8bit)	62.5	40	40.9	90.9	76.7	34
minimax-m2.5 (MLX 8bit)	50	40	52	81.8	76.7	32.8

Objective by case (%)

	recon	collection	web	privesc	lateral	cross-case
claude-opus-4-8 (thinking)	90.4	79.1	73.3	85.4	75	61.4
claude-opus-4-8	75	55.6	84.4	91.7	65	68
claude-sonnet-4-5	87.4	84.6	94.5	90.5	97.2	67.1
qwen3.7-plus	69.5	71.3	91.1	89.1	60	51.7
openai/gpt-5.6-sol	67.5	55.6	51.1	25	54.3	70.8
glm-5.2	62.1	66.7	80.1	74.7	85	67.1
qwen3.6-27b	75.4	51.3	80	76.4	85	70
google/gemma-4-31b-it	74	58.7	72.7	61.9	50	43.8
gemma-4-26b-a4b-it-heretic-guff	68.5	60.7	70.4	49	49	67.3
claude-haiku-4-5	48.4	40	80.7	70.8	88.3	53.3
minimax-m2.7-heretic (MLX 8bit)	56.4	55.6	68.3	63.4	30	55.4
minimax-m2.5 (MLX 8bit)	47	36.8	78	40.5	45	64.6

Tasks by scenario — LLM judge (/100)

	task-01	task-02	task-03	task-04	task-05
claude-opus-4-8 (thinking)	92.5	75	42	85	47.5
claude-opus-4-8	95	82.5	55	80	77
claude-sonnet-4-5	57	70	43	85	58
qwen3.7-plus	75	70	42	85	45
openai/gpt-5.6-sol	90	76	49	82.5	85
glm-5.2	68	59	57.5	0	46
qwen3.6-27b	40	60	26.5	0	55
google/gemma-4-31b-it	28	28	38	53	13
gemma-4-26b-a4b-it-heretic-guff	25	18	33	48	35
claude-haiku-4-5	29	34	36	0	34
minimax-m2.7-heretic (MLX 8bit)	24	50	6	60	16.5
minimax-m2.5 (MLX 8bit)	49	0	33	60	12.5

Judge robustness — Opus-5 vs GPT-5.6 (tasks %)

Same agent reports, two independent judges (one Claude, one non-Claude). Pearson r = 0.95 across 20 task instances. Model ranking is identical and shows no same-family favoritism — the non-Claude judge does not rank Claude higher.

	Opus-5	GPT-5.6	Δ
qwen3.7-plus	63.4	59.1	-4.3
claude-sonnet-4-5	62.6	58.1	-4.5
glm-5.2	46.1	36.5	-9.6
claude-haiku-4-5	26.6	28.1	1.5

Data table

section	item	claude-opus-4-8 (thinking)	claude-opus-4-8	claude-sonnet-4-5	qwen3.7-plus	openai/gpt-5.6-sol	glm-5.2	qwen3.6-27b	google/gemma-4-31b-it	gemma-4-26b-a4b-it-heretic-guff	claude-haiku-4-5	minimax-m2.7-heretic (MLX 8bit)	minimax-m2.5 (MLX 8bit)
Headline (%)	Objective	78.3	73.1	87.8	72.3	54.5	72.6	73.4	61.1	60.8	63.7	54.3	51.4
Headline (%)	Tasks	68.4	77.9	62.6	63.4	76.5	46.1	36.3	32	31.8	26.6	31.3	30.9
Objective by difficulty (%)	easy	90	70	100	80	70	90	70	50	60	50	40	30
Objective by difficulty (%)	medium	82.7	76.2	91.4	70.9	56.3	75	73.9	69.6	61	66.9	60.7	54.3
Objective by difficulty (%)	hard	63.3	69.5	73.8	69.8	41.6	57.6	74.6	53.1	61.2	66.8	51.9	59.7
Objective by question type (%)	boolean	75	50	100	75	50	62.5	62.5	62.5	62.5	75	62.5	50
Objective by question type (%)	extraction	90	80	90	70	60	80	70	55	50	50	40	40
Objective by question type (%)	labeling	31.1	41.8	70.1	60	35.2	26.2	69.7	37.6	54.9	47.2	40.9	52
Objective by question type (%)	mcq	100	100	100	100	63.6	100	100	90.9	100	100	90.9	81.8
Objective by question type (%)	ordering	90	91.7	85	68.3	85	66.7	78.3	41.7	75	63.3	76.7	76.7
Objective by question type (%)	set	49.2	56.6	64.9	45.7	36.9	57.2	57.1	53.5	32.7	47.4	34	32.8
Objective by case (%)	case-01-recon	90.4	75	87.4	69.5	67.5	62.1	75.4	74	68.5	48.4	56.4	47
Objective by case (%)	case-02-collection-exfil	79.1	55.6	84.6	71.3	55.6	66.7	51.3	58.7	60.7	40	55.6	36.8
Objective by case (%)	case-03-web-exploit-revshell	73.3	84.4	94.5	91.1	51.1	80.1	80	72.7	70.4	80.7	68.3	78
Objective by case (%)	case-04-privesc	85.4	91.7	90.5	89.1	25	74.7	76.4	61.9	49	70.8	63.4	40.5
Objective by case (%)	case-05-lateral-movement	75	65	97.2	60	54.3	85	85	50	49	88.3	30	45
Objective by case (%)	cross-case	61.4	68	67.1	51.7	70.8	67.1	70	43.8	67.3	53.3	55.4	64.6
Tasks by scenario — LLM judge (/100)	task-01	92.5	95	57	75	90	68	40	28	25	29	24	49
Tasks by scenario — LLM judge (/100)	task-02	75	82.5	70	70	76	59	60	28	18	34	50	0
Tasks by scenario — LLM judge (/100)	task-03	42	55	43	42	49	57.5	26.5	38	33	36	6	33
Tasks by scenario — LLM judge (/100)	task-04	85	80	85	85	82.5	0	0	53	48	0	60	60
Tasks by scenario — LLM judge (/100)	task-05	47.5	77	58	45	85	46	55	13	35	34	16.5	12.5

Heatmap cells are shaded by score (light→dark = low→high). Generated from runner/results/*.json.