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.

M1 · Evaluation awareness 2 values

M1 Evaluation awareness: values about GPT-5.4 Thinking
Evaluation and metricValueDocument
From its own documentGPT-5.4 Thinking System Card · 5 Mar 2026
Evaluation awarenessSamples with verbalized eval awarenessRun by Apollo ResearchGPT-5.4 Thinking System Card5 Mar 2026
From other documents
Evaluation awarenessSamples with moderate-or-higher eval awarenessRun by Apollo ResearchSnapshot: pre-releaseGPT-5.5 System Card23 Apr 2026First reportedGPT-5.5 System Card23 Apr 2026First reported

M3 · Sabotage and sandbagging 2 values

M3 Sabotage and sandbagging: values about GPT-5.4 Thinking
Evaluation and metricValueDocument
From its own documentGPT-5.4 Thinking System Card · 5 Mar 2026
Sabotage capabilityMean sabotage scoreRun by Apollo ResearchGPT-5.4 Thinking System Card5 Mar 2026
SandbaggingAccuracy drop when incentivized to underperformRun by Apollo ResearchGPT-5.4 Thinking System Card5 Mar 2026

M4 · Misalignment audits 2 values

M4 Misalignment audits: values about GPT-5.4 Thinking
Evaluation and metricValueDocument
From its own documentGPT-5.4 Thinking System Card · 5 Mar 2026
Covert deceptive behaviorOverall deception rateCondition: no goal nudgeRun by Apollo ResearchGPT-5.4 Thinking System Card5 Mar 2026
Destructive action avoidanceAvoidanceGPT-5.4 Thinking System Card5 Mar 2026

M5 · Honesty and hallucination 1 value

M5 Honesty and hallucination: values about GPT-5.4 Thinking
Evaluation and metricValueDocument
From other documents
Impossible Coding TaskSamples lying about completing taskRun by Apollo ResearchGPT-5.5 System Card23 Apr 2026First reportedGPT-5.5 System Card23 Apr 2026First reported

M7 · Harmful compliance and over-refusal 9 values

M7 Harmful compliance and over-refusal: values about GPT-5.4 Thinking
Evaluation and metricValueDocument
From its own documentGPT-5.4 Thinking System Card · 5 Mar 2026
Production BenchmarksExtremismGPT-5.4 Thinking System Card5 Mar 2026
Production BenchmarksHarassmentGPT-5.4 Thinking System Card5 Mar 2026
Production BenchmarksHateGPT-5.4 Thinking System Card5 Mar 2026
Production BenchmarksNonviolent illicit behaviorGPT-5.4 Thinking System Card5 Mar 2026
Production BenchmarksSelf-harm (standard)GPT-5.4 Thinking System Card5 Mar 2026
Production BenchmarksSexualGPT-5.4 Thinking System Card5 Mar 2026
Production BenchmarksSexual/minorsGPT-5.4 Thinking System Card5 Mar 2026
Production BenchmarksViolenceGPT-5.4 Thinking System Card5 Mar 2026
Production BenchmarksViolent illicit behaviorGPT-5.4 Thinking System Card5 Mar 2026

M8 · Jailbreak robustness 1 value

M8 Jailbreak robustness: values about GPT-5.4 Thinking
Evaluation and metricValueDocument
From its own documentGPT-5.4 Thinking System Card · 5 Mar 2026
JailbreaksWorst-case defender successGPT-5.4 Thinking System Card5 Mar 2026

M9 · Prompt injection 3 values

M9 Prompt injection: values about GPT-5.4 Thinking
Evaluation and metricValueDocument
From its own documentGPT-5.4 Thinking System Card · 5 Mar 2026
Prompt injectionConnectorsGPT-5.4 Thinking System Card5 Mar 2026
Prompt injectionFunction callsGPT-5.4 Thinking System Card5 Mar 2026
From other documents
Prompt injectionSearch and function-callingGPT-5.6 System Card9 Jul 2026First reportedGPT-5.6 System Card9 Jul 2026First reported

M10 · Dangerous capabilities and risk determinations 5 values

M10 Dangerous capabilities and risk determinations: values about GPT-5.4 Thinking
Evaluation and metricValueDocument
From its own documentGPT-5.4 Thinking System Card · 5 Mar 2026
Cyber RangeCombined pass rateRestated later: compare with the later valueGPT-5.4 Thinking System Card5 Mar 2026Restated later: compare with the later value
Preparedness Framework determinationAI self-improvementGPT-5.4 Thinking System Card5 Mar 2026
Preparedness Framework determinationBiological and chemicalGPT-5.4 Thinking System Card5 Mar 2026
Preparedness Framework determinationCybersecurityGPT-5.4 Thinking System Card5 Mar 2026
From other documents
Cyber RangeCombined pass rateGPT-5.5 System Card23 Apr 2026Restated: compare with the earlier valueGPT-5.5 System Card23 Apr 2026Restated: compare with the earlier value

M12 · Chain-of-thought monitorability 4 values

M12 Chain-of-thought monitorability: values about GPT-5.4 Thinking
Evaluation and metricValueDocument
From its own documentGPT-5.4 Thinking System Card · 5 Mar 2026
CoT controllabilityCoTs successfully controlledCondition: 10k-character CoTsGPT-5.4 Thinking System Card5 Mar 2026
CoT monitorabilityMonitorability vs GPT-5 ThinkingGPT-5.4 Thinking System Card5 Mar 2026
Multilingual reasoning anomaliesSamples with anomaliesRun by Apollo ResearchGPT-5.4 Thinking System Card5 Mar 2026
From other documents
CoT controllabilityCoTs successfully controlledCondition: ~5k-token CoTsGPT-5.6 System Card9 Jul 2026First reportedGPT-5.6 System Card9 Jul 2026First reported

Restated in later documents

A later document reported a value about GPT-5.4 Thinking 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 GPT-5.4 Thinking restated in later documents
Earlier valueLater value
M10 · Dangerous capabilities and risk determinations
Cyber RangeCombined pass rate
GPT-5.4 Thinking System Card5 Mar 2026 · its own documentGPT-5.5 System Card23 Apr 2026No reason stated

Risk determinations

The developer's formal decisions about GPT-5.4 Thinking under its framework, as printed. Levels from different frameworks do not map onto one another.

Risk determinations about GPT-5.4 Thinking
DomainLevel as printedFramework and document
Bio/chemhigh code, not the printed wording Preparedness Framework GPT-5.4 Thinking System Card · 5 Mar 2026Preparedness Framework GPT-5.4 Thinking System Card · 5 Mar 2026
Cyberhigh code, not the printed wording Preparedness Framework GPT-5.4 Thinking System Card · 5 Mar 2026Preparedness Framework GPT-5.4 Thinking System Card · 5 Mar 2026
AI R&D / autonomybelow_high code, not the printed wording Preparedness Framework GPT-5.4 Thinking System Card · 5 Mar 2026Preparedness Framework GPT-5.4 Thinking System Card · 5 Mar 2026

“Code, not the printed wording”: the version 0 file stored a code for this level rather than the words the document printed. The wording will be read again from the source.