Values by metric family

Each value as the document printed it. Select a value for its source, its checks and its history. Values from other documents are listed after the model’s own, each marked as the first report of that measure or as a restatement.

KeyVerifiedUnverifiedDisputedCorrected read off a figure or stated in wordsA value opens its source and history.

M2 · Reward hacking 2 values

M2 Reward hacking: values about o3
Evaluation and metricValueDocument
From its own documentOpenAI o3 and o4-mini System Card · 16 Apr 2025
RE-Bench Optimize a KernelRuns tampering with scoring functionCondition: out of 24 experimentsRun by METROpenAI o3 and o4-mini System Card16 Apr 2025
Reward hacking (all tasks)Share of task attempts with reward hackingRun by METROpenAI o3 and o4-mini System Card16 Apr 2025

M3 · Sabotage and sandbagging 2 values

M3 Sabotage and sandbagging: values about o3
Evaluation and metricValueDocument
From its own documentOpenAI o3 and o4-mini System Card · 16 Apr 2025
AI R&D sabotageAverage sabotage scoreRun by Apollo ResearchOpenAI o3 and o4-mini System Card16 Apr 2025
SandbaggingIn-context sandbagging observedCondition: including without an in-context goalRun by Apollo ResearchOpenAI o3 and o4-mini System Card16 Apr 2025

M4 · Misalignment audits 1 value

M4 Misalignment audits: values about o3
Evaluation and metricValueDocument
From other documents
Covert deceptive behaviorShare of trajectories with deceptive actionCondition: Apollo scheming suiteRun by Apollo ResearchGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported

M5 · Honesty and hallucination 8 values

M5 Honesty and hallucination: values about o3
Evaluation and metricValueDocument
From its own documentOpenAI o3 and o4-mini System Card · 16 Apr 2025
PersonQAHallucination rateCondition: no browsingOpenAI o3 and o4-mini System Card16 Apr 2025
SimpleQAHallucination rateCondition: no browsingRestated later: compare with the later valueOpenAI o3 and o4-mini System Card16 Apr 2025Restated later: compare with the later value
From other documents
AbstentionBenchRecall (appropriate abstention)GPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
Deception evalAgentic coding (impossible tasks)GPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
Deception evalBrowsing broken toolsGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
Deception evalCharXiv missing imageGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
Production CoT deception monitorShare of responses flagged deceptiveCondition: representative production trafficGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
SimpleQAHallucination rateCondition: no browsingGPT-5 System Card7 Aug 2025Restated: compare with the earlier valueGPT-5 System Card7 Aug 2025Restated: compare with the earlier value

M7 · Harmful compliance and over-refusal 13 values

M7 Harmful compliance and over-refusal: values about o3
Evaluation and metricValueDocument
From its own documentOpenAI o3 and o4-mini System Card · 16 Apr 2025
Challenging refusal evalNot unsafe (aggregate)OpenAI o3 and o4-mini System Card16 Apr 2025
Standard refusal evalNot overrefuse (aggregate)OpenAI o3 and o4-mini System Card16 Apr 2025
From other documents
Production BenchmarksExtremismGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
Production BenchmarksHarassment/threateningGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
Production BenchmarksHate/threateningGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
Production BenchmarksIllicit/non-violentGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
Production BenchmarksIllicit/violentGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
Production BenchmarksNon-violent hateGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
Production BenchmarksPersonal dataGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
Production BenchmarksSelf-harm/instructionsGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
Production BenchmarksSelf-harm/intentGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
Production BenchmarksSexual/exploitativeGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
Production BenchmarksSexual/minorsGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported

M8 · Jailbreak robustness 6 values

M8 Jailbreak robustness: values about o3
Evaluation and metricValueDocument
From its own documentOpenAI o3 and o4-mini System Card · 16 Apr 2025
Human sourced jailbreaksnot_unsafeOpenAI o3 and o4-mini System Card16 Apr 2025
StrongRejectnot_unsafe (aggregate)OpenAI o3 and o4-mini System Card16 Apr 2025
From other documents
StrongRejectAbuse/disinformation/hateGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
StrongRejectIllicit/non-violent crimeGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
StrongRejectSexual contentGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
StrongRejectViolenceGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported

M9 · Prompt injection 4 values

M9 Prompt injection: values about o3
Evaluation and metricValueDocument
From its own documentOpenAI o3 and o4-mini System Card · 16 Apr 2025
Instruction hierarchySystem<>user conflictOpenAI o3 and o4-mini System Card16 Apr 2025
From other documents
Prompt injectionBrowsingGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
Prompt injectionCodingGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported
Prompt injectionTool callingGPT-5 System Card7 Aug 2025First reportedGPT-5 System Card7 Aug 2025First reported

M10 · Dangerous capabilities and risk determinations 14 values

M10 Dangerous capabilities and risk determinations: values about o3
Evaluation and metricValueDocument
From its own documentOpenAI o3 and o4-mini System Card · 16 Apr 2025
50% time horizonTask length at 50% successRun by METRRestated later: compare with the later valueOpenAI o3 and o4-mini System Card16 Apr 2025Restated later: compare with the later value
Capture the FlagCollegiateCondition: no browsing; pass@12OpenAI o3 and o4-mini System Card16 Apr 2025
Capture the FlagHigh schoolCondition: no browsing; pass@12OpenAI o3 and o4-mini System Card16 Apr 2025
Capture the FlagProfessionalCondition: no browsing; pass@12OpenAI o3 and o4-mini System Card16 Apr 2025
Cyber RangeScenarios solved unaidedCondition: without solver code; of 2 scenariosOpenAI o3 and o4-mini System Card16 Apr 2025
OpenAI PRsPass rateOpenAI o3 and o4-mini System Card16 Apr 2025
Preparedness Framework determinationAI self-improvementOpenAI o3 and o4-mini System Card16 Apr 2025
Preparedness Framework determinationBiological and chemicalOpenAI o3 and o4-mini System Card16 Apr 2025
Preparedness Framework determinationCybersecurityOpenAI o3 and o4-mini System Card16 Apr 2025
SWE-bench VerifiedPass rateCondition: helpful-onlyOpenAI o3 and o4-mini System Card16 Apr 2025
SWE-LancerDollars earnedCondition: helpful-only; with browsingOpenAI o3 and o4-mini System Card16 Apr 2025
SWE-LancerDollars earnedCondition: no browsingOpenAI o3 and o4-mini System Card16 Apr 2025
SWE-LancerIC SWE accuracyCondition: helpful-only; with browsingOpenAI o3 and o4-mini System Card16 Apr 2025
From other documents
50% time horizonTask length at 50% successRun by METRGPT-5 System Card7 Aug 2025Restated: compare with the earlier valueGPT-5 System Card7 Aug 2025Restated: compare with the earlier value

Restated in later documents

A later document reported values about o3 again. Each pair is shown side by side: both values are kept with their own documents, and the later one does not replace the earlier.

Values about o3 restated in later documents
Earlier valueLater value
M5 · Honesty and hallucination
SimpleQAHallucination rateCondition: no browsing
OpenAI o3 and o4-mini System Card16 Apr 2025 · its own documentGPT-5 System Card7 Aug 2025No reason stated
M10 · Dangerous capabilities and risk determinations
50% time horizonTask length at 50% successRun by METR
OpenAI o3 and o4-mini System Card16 Apr 2025 · its own documentGPT-5 System Card7 Aug 2025No reason stated

Risk determinations

The developer's formal decisions about o3 under its framework, as printed. Levels from different frameworks do not map onto one another.

Risk determinations about o3
DomainLevel as printedFramework and document
Bio/chemBelow High Preparedness Framework OpenAI o3 and o4-mini System Card · 16 Apr 2025Preparedness Framework OpenAI o3 and o4-mini System Card · 16 Apr 2025
CyberBelow High Preparedness Framework OpenAI o3 and o4-mini System Card · 16 Apr 2025Preparedness Framework OpenAI o3 and o4-mini System Card · 16 Apr 2025
AI R&D / autonomyBelow High Preparedness Framework OpenAI o3 and o4-mini System Card · 16 Apr 2025Preparedness Framework OpenAI o3 and o4-mini System Card · 16 Apr 2025