GPT-5 System Card

A system card by OpenAI about gpt-5-thinking, gpt-5-main, gpt-5-thinking-mini and gpt-5-thinking-nano, published 7 Aug 2025. We recorded 118 values from it.

Developer: OpenAI
Type: System card
Models covered: gpt-5-thinking, gpt-5-main, gpt-5-thinking-mini, gpt-5-thinking-nano
Published: 7 Aug 2025
Original: The document at cdn.openai.com (opens an external site)
Also at: Another address, deploymentsafety.openai.com (opens an external site)
Archived copy: No archived copy yet
Changelog: Partial changelog

Data as of 26 Sep 2026 · Dataset v0.1 · Methodology v0.1 · Changelog

Versions

Four versions are on record, oldest first. We hold a copy of one; the others are known only by their date.

7 Aug 2025
First published version
Date from

our source registry
Copy

Known to exist; no copy held
13 Aug 2025
Revision
Date from

our source registry
Copy

Known to exist; no copy held
What changed PDF of 2025-08-07 replaced by 2025-08-13 version without changelog
24 Apr 2026
Revision
Date from

the document's changelog
Copy

Known from the changelog; no copy held
Values

2 values recorded from this version
What changed hub added chain-of-thought evaluations
26 Sep 2026
Date retrieved
Copy retrieved 26 Sep 2026
Date from

the date we retrieved it; the copy states no version date
Copy

Copy retrieved on 26 Sep 2026
Address

This version at cdn.openai.com (opens an external site)
Values

116 values recorded from this version

No version has a file hash or an archived snapshot yet. From dataset v0.2 each retrieved version carries both (Methodology §7).

Revisions

No revisions are recorded for this document. Without copies of the earlier versions, changes between them are not recorded value by value; what we know of each version is listed below.

Values

Every value we recorded from this document, grouped by metric family and ordered by where the document prints it. Location is the section, table or page as the document numbers it. 4 of the 118 have been blind-verified: a second reader found the same value without seeing ours.

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

M1 Evaluation awareness

4 values

M1 Evaluation awareness: values in the GPT-5 System Card
Model	Evaluation	Condition	Location	Checked
gpt-5-thinking	Evaluation awarenessVerbalized eval awareness, broadRun by Apollo Research	all samples	External: Apollo Research	Verified
gpt-5-thinking	Evaluation awarenessVerbalized eval awareness, broadRun by Apollo Research	scheming samples	External: Apollo Research	Unverified
gpt-5-thinking	Evaluation awarenessVerbalized eval awareness, strictRun by Apollo Research	all samples	External: Apollo Research	Unverified
gpt-5-thinking	Evaluation awarenessVerbalized eval awareness, strictRun by Apollo Research	scheming samples	External: Apollo Research	Unverified

M3 Sabotage and sandbagging

1 value

M3 Sabotage and sandbagging: values in the GPT-5 System Card
Model	Evaluation	Condition	Value	Location	Checked
gpt-5-thinking	Sandbagging reviewClear evidence of sandbaggingRun by METR	None stated		External: METR / Sandbagging	Unverified

M4 Misalignment audits

5 values

M4 Misalignment audits: values in the GPT-5 System Card
Model	Evaluation	Condition	Location	Checked
gpt-5-thinking	Covert deceptive behaviorShare of trajectories with deceptive actionRun by Apollo Research	Apollo scheming suite	External: Apollo Research	Unverified
gpt-5-thinking	Log falsificationCovert action rateRun by Apollo Research	strong goal nudge	External: Apollo Research	Unverified
gpt-5-thinking	Log falsificationCovert action rateRun by Apollo Research	weak goal nudge	External: Apollo Research	Unverified
gpt-5-thinking-helpful-only	Covert deceptive behaviorShare of trajectories with deceptive actionRun by Apollo Research	Apollo scheming suite	External: Apollo Research	Unverified
o3	Covert deceptive behaviorShare of trajectories with deceptive actionRun by Apollo Research	Apollo scheming suite	External: Apollo Research	Unverified

M5 Honesty and hallucination

21 values

M5 Honesty and hallucination: values in the GPT-5 System Card
Model	Evaluation	Condition	Location	Checked
GPT-4o	SimpleQAHallucination rate	no browsing	3.7 Table 8	Unverified
gpt-5-main	SimpleQAHallucination rate	no browsing	3.7 Table 8	Unverified
gpt-5-thinking	SimpleQAHallucination rate	no browsing	3.7 Table 8	Unverified
gpt-5-thinking-mini	SimpleQAHallucination rate	no browsing	3.7 Table 8	Unverified
gpt-5-thinking-nano	SimpleQAHallucination rate	no browsing	3.7 Table 8	Unverified
o3	SimpleQAHallucination rate	no browsing	3.7 Table 8	Unverified
o4-mini	SimpleQAHallucination rate	no browsing	3.7 Table 8	Unverified
gpt-5-main	Production-traffic factualityClaim-level hallucination rate reduction vs GPT-4o	with browsing; production prompts	3.7 text	Unverified
gpt-5-main	Production-traffic factualityReduction in responses with 1+ major error vs GPT-4o	with browsing; production prompts	3.7 text	Unverified
gpt-5-thinking	Production-traffic factualityClaim-level hallucination rate reduction vs o3	with browsing; production prompts	3.7 text	Unverified
gpt-5-thinking	Production-traffic factualityReduction in responses with 1+ major error vs o3	with browsing; production prompts	3.7 text	Unverified
gpt-5-thinking	AbstentionBenchRecall (appropriate abstention)	None stated	3.8 Table 9	Unverified
gpt-5-thinking	Deception evalAgentic coding (impossible tasks)	None stated	3.8 Table 9	Unverified
gpt-5-thinking	Deception evalBrowsing broken tools	None stated	3.8 Table 9	Unverified
gpt-5-thinking	Deception evalCharXiv missing image	None stated	3.8 Table 9	Unverified
o3	AbstentionBenchRecall (appropriate abstention)	None stated	3.8 Table 9	Unverified
o3	Deception evalAgentic coding (impossible tasks)	None stated	3.8 Table 9	Unverified
o3	Deception evalBrowsing broken tools	None stated	3.8 Table 9	Unverified
o3	Deception evalCharXiv missing image	None stated	3.8 Table 9	Unverified
gpt-5-thinking	Production CoT deception monitorShare of responses flagged deceptive	representative production traffic	3.8.1	Verified
o3	Production CoT deception monitorShare of responses flagged deceptive	representative production traffic	3.8.1	Verified

M6 Sycophancy

5 values

M6 Sycophancy: values in the GPT-5 System Card
Model	Evaluation	Condition	Location	Checked
GPT-4o	Sycophancy offline evalSycophancy score	offline	3.3 Table 4	Unverified
gpt-5-main	Sycophancy offline evalSycophancy score	offline	3.3 Table 4	Unverified
gpt-5-main	Sycophancy online prevalence (A/B)Relative change vs GPT-4o	free users	3.3 Table 4	Unverified
gpt-5-main	Sycophancy online prevalence (A/B)Relative change vs GPT-4o	paid users	3.3 Table 4	Unverified
gpt-5-thinking	Sycophancy offline evalSycophancy score	offline	3.3 Table 4	Unverified

M7 Harmful compliance and over-refusal

44 values

M7 Harmful compliance and over-refusal: values in the GPT-5 System Card
Model	Evaluation	Condition	Location	Checked
GPT-4o	Production BenchmarksExtremism	None stated	3.2 Table 3	Unverified
GPT-4o	Production BenchmarksHarassment/threatening	None stated	3.2 Table 3	Unverified
GPT-4o	Production BenchmarksHate/threatening	None stated	3.2 Table 3	Unverified
GPT-4o	Production BenchmarksIllicit/non-violent	None stated	3.2 Table 3	Unverified
GPT-4o	Production BenchmarksIllicit/violent	None stated	3.2 Table 3	Unverified
GPT-4o	Production BenchmarksNon-violent hate	None stated	3.2 Table 3	Unverified
GPT-4o	Production BenchmarksPersonal data	None stated	3.2 Table 3	Unverified
GPT-4o	Production BenchmarksSelf-harm/instructions	None stated	3.2 Table 3	Unverified
GPT-4o	Production BenchmarksSelf-harm/intent	None stated	3.2 Table 3	Unverified
GPT-4o	Production BenchmarksSexual/exploitative	None stated	3.2 Table 3	Unverified
GPT-4o	Production BenchmarksSexual/minors	None stated	3.2 Table 3	Unverified
gpt-5-main	Production BenchmarksExtremism	None stated	3.2 Table 3	Unverified
gpt-5-main	Production BenchmarksHarassment/threatening	None stated	3.2 Table 3	Unverified
gpt-5-main	Production BenchmarksHate/threatening	None stated	3.2 Table 3	Unverified
gpt-5-main	Production BenchmarksIllicit/non-violent	None stated	3.2 Table 3	Unverified
gpt-5-main	Production BenchmarksIllicit/violent	None stated	3.2 Table 3	Unverified
gpt-5-main	Production BenchmarksNon-violent hate	None stated	3.2 Table 3	Unverified
gpt-5-main	Production BenchmarksPersonal data	None stated	3.2 Table 3	Unverified
gpt-5-main	Production BenchmarksSelf-harm/instructions	None stated	3.2 Table 3	Unverified
gpt-5-main	Production BenchmarksSelf-harm/intent	None stated	3.2 Table 3	Unverified
gpt-5-main	Production BenchmarksSexual/exploitative	None stated	3.2 Table 3	Unverified
gpt-5-main	Production BenchmarksSexual/minors	None stated	3.2 Table 3	Unverified
gpt-5-thinking	Production BenchmarksExtremism	None stated	3.2 Table 3	Unverified
gpt-5-thinking	Production BenchmarksHarassment/threatening	None stated	3.2 Table 3	Unverified
gpt-5-thinking	Production BenchmarksHate/threatening	None stated	3.2 Table 3	Verified
gpt-5-thinking	Production BenchmarksIllicit/non-violent	None stated	3.2 Table 3	Unverified
gpt-5-thinking	Production BenchmarksIllicit/violent	None stated	3.2 Table 3	Unverified
gpt-5-thinking	Production BenchmarksNon-violent hate	None stated	3.2 Table 3	Unverified
gpt-5-thinking	Production BenchmarksPersonal data	None stated	3.2 Table 3	Unverified
gpt-5-thinking	Production BenchmarksSelf-harm/instructions	None stated	3.2 Table 3	Unverified
gpt-5-thinking	Production BenchmarksSelf-harm/intent	None stated	3.2 Table 3	Unverified
gpt-5-thinking	Production BenchmarksSexual/exploitative	None stated	3.2 Table 3	Unverified
gpt-5-thinking	Production BenchmarksSexual/minors	None stated	3.2 Table 3	Unverified
o3	Production BenchmarksExtremism	None stated	3.2 Table 3	Unverified
o3	Production BenchmarksHarassment/threatening	None stated	3.2 Table 3	Unverified
o3	Production BenchmarksHate/threatening	None stated	3.2 Table 3	Unverified
o3	Production BenchmarksIllicit/non-violent	None stated	3.2 Table 3	Unverified
o3	Production BenchmarksIllicit/violent	None stated	3.2 Table 3	Unverified
o3	Production BenchmarksNon-violent hate	None stated	3.2 Table 3	Unverified
o3	Production BenchmarksPersonal data	None stated	3.2 Table 3	Unverified
o3	Production BenchmarksSelf-harm/instructions	None stated	3.2 Table 3	Unverified
o3	Production BenchmarksSelf-harm/intent	None stated	3.2 Table 3	Unverified
o3	Production BenchmarksSexual/exploitative	None stated	3.2 Table 3	Unverified
o3	Production BenchmarksSexual/minors	None stated	3.2 Table 3	Unverified

M8 Jailbreak robustness

16 values

M8 Jailbreak robustness: values in the GPT-5 System Card
Model	Evaluation	Condition	Location	Checked
GPT-4o	StrongRejectAbuse/disinformation/hate	None stated	3.4 Table 5	Unverified
GPT-4o	StrongRejectIllicit/non-violent crime	None stated	3.4 Table 5	Unverified
GPT-4o	StrongRejectSexual content	None stated	3.4 Table 5	Unverified
GPT-4o	StrongRejectViolence	None stated	3.4 Table 5	Unverified
gpt-5-main	StrongRejectAbuse/disinformation/hate	None stated	3.4 Table 5	Unverified
gpt-5-main	StrongRejectIllicit/non-violent crime	None stated	3.4 Table 5	Unverified
gpt-5-main	StrongRejectSexual content	None stated	3.4 Table 5	Unverified
gpt-5-main	StrongRejectViolence	None stated	3.4 Table 5	Unverified
gpt-5-thinking	StrongRejectAbuse/disinformation/hate	None stated	3.4 Table 5	Unverified
gpt-5-thinking	StrongRejectIllicit/non-violent crime	None stated	3.4 Table 5	Unverified
gpt-5-thinking	StrongRejectSexual content	None stated	3.4 Table 5	Unverified
gpt-5-thinking	StrongRejectViolence	None stated	3.4 Table 5	Unverified
o3	StrongRejectAbuse/disinformation/hate	None stated	3.4 Table 5	Unverified
o3	StrongRejectIllicit/non-violent crime	None stated	3.4 Table 5	Unverified
o3	StrongRejectSexual content	None stated	3.4 Table 5	Unverified
o3	StrongRejectViolence	None stated	3.4 Table 5	Unverified

M9 Prompt injection

6 values

M9 Prompt injection: values in the GPT-5 System Card
Model	Evaluation	Condition	Location	Checked
gpt-5-thinking	Prompt injectionBrowsing	None stated	3.6 Table 7	Unverified
gpt-5-thinking	Prompt injectionCoding	None stated	3.6 Table 7	Unverified
gpt-5-thinking	Prompt injectionTool calling	None stated	3.6 Table 7	Unverified
o3	Prompt injectionBrowsing	None stated	3.6 Table 7	Unverified
o3	Prompt injectionCoding	None stated	3.6 Table 7	Unverified
o3	Prompt injectionTool calling	None stated	3.6 Table 7	Unverified

M10 Dangerous capabilities and risk determinations

14 values

M10 Dangerous capabilities and risk determinations: values in the GPT-5 System Card
Model	Evaluation	Condition	Location	Checked
ChatGPT agent	MLE-BenchMedal rate	None stated	5 AI Self-improvement	Unverified
gpt-5-thinking	OpenAI-Proof Q&APass rate	None stated	5 AI Self-improvement	Unverified
gpt-5-thinking	TroubleshootingBenchDifference vs o3	None stated	5 Bio	Unverified
gpt-5-thinking	Preparedness Framework determinationAI self-improvement	None stated	5 Preparedness	Unverified
gpt-5-thinking	Preparedness Framework determinationBiological and chemical	None stated	5 Preparedness	Unverified
gpt-5-thinking	Preparedness Framework determinationCybersecurity	None stated	5 Preparedness	Unverified
gpt-5-thinking	50% time horizonTask length at 50% successRun by METR	None stated	External: METR	Unverified
o3	50% time horizonTask length at 50% successRun by METR	None stated	External: METR	Unverified
gpt-5-thinking	Cyber challengesEasy: challenges solvedRun by Irregular	out of 18 challenges	External: Pattern Labs	Unverified
gpt-5-thinking	Cyber challengesEvasion: average success rateRun by Irregular	None stated	External: Pattern Labs	Unverified
gpt-5-thinking	Cyber challengesHard: challenges solvedRun by Irregular	out of 4 challenges	External: Pattern Labs	Unverified
gpt-5-thinking	Cyber challengesMedium: challenges solvedRun by Irregular	out of 14 challenges	External: Pattern Labs	Unverified
gpt-5-thinking	Cyber challengesNetwork attack simulation: average success rateRun by Irregular	None stated	External: Pattern Labs	Unverified
gpt-5-thinking	Cyber challengesVulnerability discovery and exploitation: average success rateRun by Irregular	None stated	External: Pattern Labs	Unverified

M12 Chain-of-thought monitorability

2 values

M12 Chain-of-thought monitorability: values in the GPT-5 System Card
Model	Evaluation	Condition	Value	Location	Checked	Version
gpt-5-thinking	CoT controllabilityAbility to control CoT content	None stated		CoT Evaluations (hub)	Unverified	Version of 24 Apr 2026
gpt-5-thinking	CoT monitorabilityMonitorability vs o3	None stated		CoT Evaluations (hub)	Unverified	Version of 24 Apr 2026

Extraction coverage

What we read of this document, and where each value was read.

Our note Many Preparedness numbers are chart-only (excluded)

All 118 values were read in the document itself.

Of the 118 values, 4 are statements in words rather than numbers; they are marked * and left out of charts by default.

All 118 values were extracted for version 0 of the dataset through a web reader, which did not always reach the later sections of long PDFs (Methodology §3).

Where next

OpenAIEvery OpenAI document, model and changelog practice
gpt-5-thinkingEvery value about this model, from every document
gpt-5-mainEvery value about this model, from every document
gpt-5-thinking-miniEvery value about this model, from every document
gpt-5-thinking-nanoEvery value about this model, from every document
All documentsEvery document by developer and date
Revisions ledgerEvery recorded change to a document, newest first
M1 · Evaluation awarenessThis family across every developer and document
M3 · Sabotage and sandbaggingThis family across every developer and document
M4 · Misalignment auditsThis family across every developer and document
M5 · Honesty and hallucinationThis family across every developer and document
M6 · SycophancyThis family across every developer and document
M7 · Harmful compliance and over-refusalThis family across every developer and document
M8 · Jailbreak robustnessThis family across every developer and document
M9 · Prompt injectionThis family across every developer and document
M10 · Dangerous capabilities and risk determinationsThis family across every developer and document
M12 · Chain-of-thought monitorabilityThis family across every developer and document