Examenos

49.6% · raw 49.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation; highest reasoning budget tested, not mapped to an effort tier. Mean rubric score on 99 problems, up to nine scored responses per problem, at the highest reasoning budget tested. Evaluation date unpublished.

37.7% · raw 37.7 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation; highest reasoning budget tested, not mapped to an effort tier. Single-attempt success estimated from seven responses on each of 129 problems, at the highest reasoning budget tested; benchmark-specific configuration differs from production. Evaluation date unpublished.

61.3% · raw 61.3 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation; highest reasoning budget tested, not mapped to an effort tier. Research tasks from practising life scientists; 94 problems after excluding six flagged ones, up to eight scored responses per problem, at the highest reasoning budget tested. Evaluation date unpublished.

42% · raw 42 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

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 Evaluation date unpublished.

42% · raw 42 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

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 Evaluation date unpublished.

59% · raw 59 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

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 Evaluation date unpublished.

59% · raw 59 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

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 Evaluation date unpublished.

78.3% · raw 78.3 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

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 Evaluation date unpublished.

85% · raw 85 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

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

59.9% · raw 59.9 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal biological capability evaluation; high capability evaluation. Higher means more of the tested capability, not more safety. Evaluation date unpublished.

59.9% · raw 59.9 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal biological capability evaluation; high capability evaluation. Higher means more of the tested capability, not more safety. Evaluation date unpublished.

9.4% · raw 9.4 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal biological capability evaluation. Evaluation date unpublished.

0.5 Spearman · raw 0.529 Spearman

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal biological capability evaluation; Critical capability evaluation. Evaluation date unpublished.

0.4 R² · raw 0.38 R²

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

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

0.4 score · raw 0.42 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

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

13.4 NLL · raw 13.41 NLL

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal biological capability evaluation; indicative Critical threshold is an NLL of 9.40 or lower. Evaluation date unpublished.

98.1% · raw 98.1 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal refusal evaluation. Fraction of responses passing the safety grader. Evaluation date unpublished.

79% · raw 79 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal refusal evaluation. Fraction of responses passing the safety grader. Evaluation date unpublished.

83.8% · raw 83.8 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal refusal evaluation. Fraction of responses passing the safety grader. Evaluation date unpublished.

84.6% · raw 84.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal refusal evaluation. Fraction of responses passing the safety grader. Evaluation date unpublished.

100% · raw 100 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal refusal evaluation. Fraction of responses passing the safety grader. Evaluation date unpublished.

100% · raw 100 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal refusal evaluation. Compliance on benign biosecurity prompts. Evaluation date unpublished.

14.3% · raw 14.3 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation. Task success under historical cyber jailbreak attacks with real-time blocking disabled. Evaluation date unpublished.

9.5% · raw 9.5 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation. Blocker-related deception rate. Evaluation date unpublished.

1.3% · raw 1.3 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation. Rate of failing to acknowledge that the search tool is unavailable. Evaluation date unpublished.

0% · raw 0 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation. Rate of successfully exploiting evaluation infrastructure outside the assigned task. Evaluation date unpublished.

21.7% · raw 21.7 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation. Rate of successfully circumventing a restriction after encountering a warning. Evaluation date unpublished.

0% · raw 0 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation. Rate of attempting to circumvent an automatic-review denial, including unsuccessful retries. Evaluation date unpublished.

27.1% · raw 27.1 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation. Rate of harmful or unauthorized actions at or above a set severity in replayed agentic conversations, excluding laziness. Evaluation date unpublished.

2.5% · raw 2.5 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation. Rate of performing the specified unauthorized action, conditional on discovering the message board. Evaluation date unpublished.

1.4% · raw 1.4 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation. Overall rate of misaligned outcomes in simulated workplace tasks. Evaluation date unpublished.

0% · raw 0 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation. Full-context monitor bypass rate among successful hidden mathematics tasks. Evaluation date unpublished.

0% · raw 0 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation. Attempts that successfully sabotage simulated services and evade the monitor, unguided condition. Evaluation date unpublished.

0% · raw 0 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation. Attempts that successfully sabotage simulated services and evade the monitor, guided condition. Evaluation date unpublished.

71.8% · raw 71.8 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation. Rate of following instructions that constrain the content or form of the reasoning, on GPQA items. Evaluation date unpublished.

59.3% · raw 59.3 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation. Rate of following instructions that constrain the content or form of the reasoning, on HLE items. Evaluation date unpublished.

83.1% · raw 83.1 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-30

Harness: OpenAI internal evaluation. Rate of following instructions that constrain the content or form of the reasoning, on MMLU items. Evaluation date unpublished.