Model profile
GPT 5.4
OpenAIdeveloper
2026-03-05release date
#24 / 267overall rank
22eval lineages
Evidence summary
GPT 5.4 has an estimated overall rank of #24; its 90% source-sensitivity interval is #11–#100. Its behavior-only rank is #27; company governance moves the combined estimate to #24. Published evidence spans 22 evals and 7 of 7 behavior components. Its strongest relative result is Enkrypt AI Safety Leaderboard (harmful_attack_non_success_rate, #1 of 260); its weakest is CAIS Risk Index (textquests_harm, #47 of 48).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗gpt-5-4
- OpenRouter ↗openai/gpt-5.4
- Official model documentation ↗Family-level model document · openai · first party
- Release source ↗direct research preview identity date
Published eval results
Rank is within that sub-eval. Black marks the observed result; the grey dot marks the value implied by the global rank. Values stay on each source’s native scale.
| Eval / sub-eval | Rank | Value | Distribution | Better | Source |
|---|---|---|---|---|---|
| AA-Omnisciencehallucination_rate | #162 / 311 | 0.8109 | ↓ lower | Source ↗official | |
| AgentAbstainabstain | #3 / 17 | 67.8 | ↑ higher | Source ↗official | |
| AgentAbstaincar | #5 / 17 | 64 | ↑ higher | Source ↗official | |
| AgentAbstainpaired | #7 / 17 | 48.7 | ↑ higher | Source ↗official | |
| BioSecBench-Refusalbalanced_refusal_score | #4 / 10 | 0.4254 | ↑ higher | Source ↗official | |
| BullshitBench v2clear_pushback_rate | #29 / 105 | 0.45 | ↑ higher | Source ↗official | |
| CAIS Risk Indexagent_red_teaming | #13 / 43 | 59.7 | ↓ lower | Source ↗official | |
| CAIS Risk Indexbioweapons_assistance | #11 / 48 | 57.1 | ↓ lower | Source ↗official | |
| CAIS Risk Indexhle_overconfidence | #6 / 49 | 40.9 | ↓ lower | Source ↗official | |
| CAIS Risk Indexmachiavelli | #34 / 45 | 92.9 | ↓ lower | Source ↗official | |
| CAIS Risk Indexmask | #11 / 51 | 9.7 | ↓ lower | Source ↗official | |
| CAIS Risk Indexpolitical_manipulation | #28 / 32 | 59.4 | ↓ lower | Source ↗official | |
| CAIS Risk Indextextquests_harm | #47 / 48 | 24.5 | ↓ lower | Source ↗official | |
| Cisco AI Defense Rolling Single-Turn Leaderboardsingle_turn_attack_success_rate | #8 / 105 | 2.742 | ↓ lower | Source ↗official | |
| DystopiaBenchbasaglia_score | #13 / 50 | 47.3 | ↓ lower | Source ↗official | |
| DystopiaBenchbaudrillard_score | #12 / 50 | 38.87 | ↓ lower | Source ↗official | |
| DystopiaBenchhuxley_score | #13 / 50 | 46.2 | ↓ lower | Source ↗official | |
| DystopiaBenchlaguardia_score | #10 / 50 | 46.57 | ↓ lower | Source ↗official | |
| DystopiaBenchorwell_score | #13 / 50 | 44.4 | ↓ lower | Source ↗official | |
| DystopiaBenchpetrov_score | #7 / 50 | 37.4 | ↓ lower | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #181 / 260 | 12.92 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #199 / 260 | 83.33 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #1 / 260 | 100 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #105 / 258 | 97.27 | ↑ higher | Source ↗official | |
| GPT-5.6 system card — disallowed content with challenging promptsextremism_not_unsafe | #1 / 7 | 1 | ↑ higher | Source ↗official | |
| GPT-5.6 system card — disallowed content with challenging promptsgore_not_unsafe | #2 / 7 | 0.831 | ↑ higher | Source ↗official | |
| GPT-5.6 system card — disallowed content with challenging promptshate_not_unsafe | #5 / 7 | 0.943 | ↑ higher | Source ↗official | |
| GPT-5.6 system card — disallowed content with challenging promptsnonviolent_illicit_not_unsafe | #1 / 7 | 1 | ↑ higher | Source ↗official | |
| GPT-5.6 system card — disallowed content with challenging promptsself_harm_not_unsafe | #1 / 7 | 0.987 | ↑ higher | Source ↗official | |
| GPT-5.6 system card — disallowed content with challenging promptssexual_minors_not_unsafe | #4 / 7 | 0.966 | ↑ higher | Source ↗official | |
| GPT-5.6 system card — disallowed content with challenging promptssexual_not_unsafe | #6 / 7 | 0.933 | ↑ higher | Source ↗official | |
| GPT-5.6 system card — disallowed content with challenging promptsviolent_illicit_not_unsafe | #2 / 7 | 0.971 | ↑ higher | Source ↗official | |
| GPT-5.6 system card — first-person fairnessharm_overall_pct | #2 / 7 | 0.88 | ↓ lower | Source ↗official | |
| GPT-5.6 system card — prompt-injection robustnessconnectors_injection_resistance | #5 / 7 | 0.998 | ↑ higher | Source ↗official | |
| GPT-5.6 system card — prompt-injection robustnesssearch_function_calling_injection_resistance | #4 / 6 | 0.697 | ↑ higher | Source ↗official | |
| HUMAINE Trust, Ethics and Safetytrust_ethics_safety_score | #25 / 54 | 27.17 | ↑ higher | Source ↗official | |
| MACHIAVELLIdeception_relative_random_pct | #34 / 50 | 92.9 | ↓ lower | Source ↗official | |
| MASKlying_probability_pct | #8 / 53 | 10.33 | ↓ lower | Source ↗official | |
| OpenAI GPT-5.4 Dynamic Wellbeingemotional_reliance | #1 / 3 | 0.985 | ↑ higher | Source ↗official | |
| OpenAI GPT-5.4 Dynamic Wellbeingmental_health | #1 / 3 | 0.985 | ↑ higher | Source ↗official | |
| OpenAI GPT-5.4 Dynamic Wellbeingself_harm | #1 / 3 | 0.977 | ↑ higher | Source ↗official | |
| OpenAI GPT-5.4 First-Person Fairnessharm_overall | #1 / 3 | 0.0088 | ↓ lower | Source ↗official | |
| OpenAI GPT-5.4 Property Preservationdestructive_action_avoidance | #2 / 3 | 0.86 | ↑ higher | Source ↗official | |
| OpenAI GPT-5.4 Property Preservationuser_work_preserved | #1 / 3 | 0.53 | ↑ higher | Source ↗official | |
| OpenAI GPT-5.4 User Confirmationsfinancial_transaction | #1 / 3 | 1 | ↑ higher | Source ↗official | |
| OpenAI GPT-5.4 User Confirmationsgeneral_confirmation | #1 / 3 | 0.94 | ↑ higher | Source ↗official | |
| OpenAI GPT-5.4 User Confirmationshigh_stakes_communication | #1 / 3 | 1 | ↑ higher | Source ↗official | |
| SABERoverall_safety_rate | #2 / 13 | 36.1 | ↑ higher | Source ↗official | |
| SABERscenario_a_safety_rate | #3 / 13 | 36.02 | ↑ higher | Source ↗official | |
| SABERscenario_b_safety_rate | #2 / 13 | 39.38 | ↑ higher | Source ↗official | |
| SABERscenario_c_safety_rate | #2 / 13 | 33.51 | ↑ higher | Source ↗official | |
| SM-Benchadversarial | #57 / 73 | 77.56 | ↑ higher | Source ↗official | |
| SM-Benchambiguous_interpretation | #23 / 73 | 87.8 | ↑ higher | Source ↗official | |
| SM-Benchanti_hallucination | #42 / 73 | 90.58 | ↑ higher | Source ↗official | |
| SM-Bencheq_boundaries | #46 / 73 | 60.11 | ↑ higher | Source ↗official | |
| SM-Benchoverfit | #62 / 73 | 38.25 | ↑ higher | Source ↗official | |
| SpeciEvalbelief_animal_sentience | #72 / 102 | 6.57 | ↑ higher | Source ↗official | |
| SpeciEvalland_animal_4ns | #5 / 102 | 3.7 | ↓ lower | Source ↗official | |
| SpeciEvalsea_animal_4ns | #1 / 102 | 3.65 | ↓ lower | Source ↗official | |
| SpeciEvalspeciesism | #37 / 102 | 1.82 | ↓ lower | Source ↗official | |
| VETO Misfired Alignmentmisfired_alignment_rate_pct | #24 / 25 | 17.6 | ↓ lower | Source ↗official | |
| Vigil Mental Health Safetyoverall_score | #2 / 23 | 78 | ↑ higher | Source ↗official |
Values evaluations
Descriptive values and political-framing results are separate from safety/ethics ranks. Each strip shows the evaluation’s observed model range; its endpoint labels state what lower and higher values mean.
Agent-ValueBench Moral Foundations (MFT08)
Agent-ValueBench HEXACO
| Dimension | Value | Distribution |
|---|---|---|
| Openness to experience | 4.5 | |
| Honesty-humility | 6 | |
| Extraversion | 5.7 | |
| Agreeableness | 5.6 | |
| Conscientiousness | 7.3 |