Examenos

96.9% · raw 96.9 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Production benchmark safe-response rate; higher is better. Raw value 0.969 on a 0–1 scale. Evaluation date unpublished.

99.7% · raw 99.7 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Production benchmark safe-response rate; higher is better. Raw value 0.997 on a 0–1 scale. Evaluation date unpublished.

89.4% · raw 89.4 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Production benchmark safe-response rate; higher is better. Raw value 0.894 on a 0–1 scale. Evaluation date unpublished.

100% · raw 100 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Production benchmark safe-response rate; higher is better. Raw value 1.000 on a 0–1 scale. Evaluation date unpublished.

91.9% · raw 91.9 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Production benchmark safe-response rate; higher is better. Raw value 0.919 on a 0–1 scale. Evaluation date unpublished.

85.8% · raw 85.8 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Production benchmark safe-response rate; higher is better. Raw value 0.858 on a 0–1 scale. Evaluation date unpublished.

89.6% · raw 89.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Production benchmark safe-response rate; higher is better. Raw value 0.896 on a 0–1 scale. Evaluation date unpublished.

88.8% · raw 88.8 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Production benchmark safe-response rate; higher is better. Raw value 0.888 on a 0–1 scale. Evaluation date unpublished.

82.4% · raw 82.4 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. U18 safe-response rate; higher is better. Raw value 0.824 on a 0–1 scale. Evaluation date unpublished.

93.3% · raw 93.3 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. U18 safe-response rate; higher is better. Raw value 0.933 on a 0–1 scale. Evaluation date unpublished.

81.7% · raw 81.7 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. U18 safe-response rate; higher is better. Raw value 0.817 on a 0–1 scale. Evaluation date unpublished.

87% · raw 87 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. U18 safe-response rate; higher is better. Raw value 0.870 on a 0–1 scale. Evaluation date unpublished.

97.9% · raw 97.9 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. U18 safe-response rate; higher is better. Raw value 0.979 on a 0–1 scale. Evaluation date unpublished.

85.8% · raw 85.8 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. U18 safe-response rate; higher is better. Raw value 0.858 on a 0–1 scale. Evaluation date unpublished.

97.5% · raw 97.5 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Vision safe-response rate; higher is better. Raw value 0.975 on a 0–1 scale. Evaluation date unpublished.

84.9% · raw 84.9 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Vision safe-response rate; higher is better. Raw value 0.849 on a 0–1 scale. Evaluation date unpublished.

97.6% · raw 97.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Vision safe-response rate; higher is better. Raw value 0.976 on a 0–1 scale. Evaluation date unpublished.

95.7% · raw 95.7 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Vision safe-response rate; higher is better. Raw value 0.957 on a 0–1 scale. Evaluation date unpublished.

97.1% · raw 97.1 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Dynamic multi-turn safe-response rate; higher is better. Raw value 0.971 on a 0–1 scale. Evaluation date unpublished.

96.4% · raw 96.4 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Dynamic multi-turn safe-response rate; higher is better. Raw value 0.964 on a 0–1 scale. Evaluation date unpublished.

92.6% · raw 92.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Dynamic multi-turn safe-response rate; higher is better. Raw value 0.926 on a 0–1 scale. Evaluation date unpublished.

97.6% · raw 97.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Biology refusal evaluation, Safe metric; higher is better. Raw value 0.976 on a 0–1 scale. Evaluation date unpublished.

92.2% · raw 92.2 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Biology refusal evaluation, Safe metric; higher is better. Raw value 0.922 on a 0–1 scale. Evaluation date unpublished.

97.6% · raw 97.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Cyber refusal safe-response rate; higher is better. Raw value 0.976 on a 0–1 scale. Evaluation date unpublished.

99.7% · raw 99.7 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Cyber refusal safe-response rate; higher is better. Raw value 0.997 on a 0–1 scale. Evaluation date unpublished.

94.5% · raw 94.5 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI GPT-Red automated evaluation, indirect attacks, averaged equally across available reasoning levels. Evaluation date unpublished.

99.9% · raw 99.91 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI GPT-Red automated evaluation, direct attacks. Stated in addendum prose. Evaluation date unpublished.

47.4% · raw 47.4 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

length-adjusted (raw 56.2, mean length 4,969 chars); GPT-6 Luna (October 29) addendum, Table 5

48.2% · raw 48.2 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

length-adjusted (raw 54.0, mean length 2,970 chars); GPT-6 Luna (October 29) addendum, Table 5

22.8% · raw 22.8 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

length-adjusted (raw 28.5, mean length 2,725 chars); GPT-6 Luna (October 29) addendum, Table 5

93.3% · raw 93.3 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

length-adjusted (raw 93.7, mean length 2,978 chars); GPT-6 Luna (October 29) addendum, Table 5

48.7% · raw 48.68 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Task-clipped metric, mean task-level score ± 0.97 one standard error, four responses per task; GPT-6 Luna (October 29) addendum, Table 6

46.4% · raw 46.39 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Task-clipped metric, mean task-level score ± 1.29 one standard error, four responses per task; GPT-6 Luna (October 29) addendum, Table 6

53.9% · raw 53.92 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Task-clipped metric, mean task-level score ± 2.28 one standard error, four responses per task; GPT-6 Luna (October 29) addendum, Table 6

49.6% · raw 49.64 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Task-clipped metric, mean task-level score ± 1.87 one standard error, four responses per task; GPT-6 Luna (October 29) addendum, Table 6

45.9% · raw 45.85 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal biological capability evaluation; pass@1, indicative threshold 31%. Higher means more of the tested capability, not more safety. Evaluation date unpublished.

31.8% · raw 31.8 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal biological capability evaluation; pass@1, indicative threshold 54%. Higher means more of the tested capability, not more safety. Evaluation date unpublished.

90.8% · raw 90.8 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal biological capability evaluation; cons@32, indicative threshold 80%. Higher means more of the tested capability, not more safety. Evaluation date unpublished.

37.9% · raw 37.9 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal biological capability evaluation; pass@1, indicative threshold 36.4%. Higher means more of the tested capability, not more safety. Evaluation date unpublished.

0.5 Spearman · raw 0.53 Spearman

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal biological capability evaluation; indicative threshold 0.60. Evaluation date unpublished.

0.2 R² · raw 0.21 R²

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal biological capability evaluation; indicative threshold 0.60. Evaluation date unpublished.

0.4 score · raw 0.37 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal biological capability evaluation; indicative threshold 0.75. Evaluation date unpublished.

11.8 NLL · raw 11.84 NLL

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal biological capability evaluation; indicative threshold 9.40 or lower. Evaluation date unpublished.

46.2% · raw 46.19 %

Headline · max effort · Own vendor

Source/record date: 2026-10-29

GPT-6 Luna (October 29) addendum §8.1.2.1, at maximum reasoning effort

51.8% · raw 51.75 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

GPT-6 Luna (October 29) addendum §8.1.2.2

11.8% · raw 11.8 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Conditional task completion after a warning; lower is better. Labelled October 29 value read visually from the addendum figure on 1 November 2026. Evaluation date unpublished.

7.8% · raw 7.83 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Rate of failing to acknowledge the unavailable search tool; lower is better. Labelled October 29 value read visually from the addendum figure on 1 November 2026. Evaluation date unpublished.

11.1% · raw 11.1 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Blocker-related deception rate; lower is better. Labelled October 29 value read visually from the addendum figure on 1 November 2026. Evaluation date unpublished.

0% · raw 0 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Prohibited actions completed after rejection; lower is better. Labelled October 29 value read visually from the addendum figure on 1 November 2026. Evaluation date unpublished.

8.5% · raw 8.5 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Responses with an error on factuality-heavy prompts; lower is better. Labelled October 29 value read visually from the addendum figure on 1 November 2026. Evaluation date unpublished.

2.5% · raw 2.5 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Claims with an error on factuality-heavy prompts; lower is better. Labelled October 29 value read visually from the addendum figure on 1 November 2026. Evaluation date unpublished.

24.6% · raw 24.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Responses with an error on prompts from user-flagged failures; lower is better. Labelled October 29 value read visually from the addendum figure on 1 November 2026. Evaluation date unpublished.

7.2% · raw 7.2 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Claims with an error on prompts from user-flagged failures; lower is better. Labelled October 29 value read visually from the addendum figure on 1 November 2026. Evaluation date unpublished.

7.8% · raw 7.8 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Responses with an error on high-stakes prompts; lower is better. Labelled October 29 value read visually from the addendum figure on 1 November 2026. Evaluation date unpublished.

1.6% · raw 1.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-29

Harness: OpenAI internal evaluation. Claims with an error on high-stakes prompts; lower is better. Labelled October 29 value read visually from the addendum figure on 1 November 2026. Evaluation date unpublished.