Model profile
GPT 5.4 Mini
Evidence summary
Safety. GPT 5.4 Mini has an estimated Safety rank of #75; its 90% source-sensitivity interval is #32–#224. Its behavior-only rank is #86; company governance moves the combined estimate to #75. Published Safety evidence spans 23 eval lineages and 6 of 7 components. Its strongest relative result is Arena Factuality — Text Arena (factuality-only weighting) (factuality_bt_rating, #1 of 111); its weakest is SM-Bench (overfit, #90 of 92).
Freedom. GPT 5.4 Mini has an estimated Freedom rank of #642; its 90% source-sensitivity interval is #473–#652. Published Freedom evidence spans 7 eval lineages and 1 of 1 components. Its strongest relative result is SM-Bench (adversarial, #14 of 92); its weakest is Adversarial Humanities Benchmark (AHB) — Table 5 (intellectual_property, #31 of 31).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗gpt-5-4-mini
- OpenRouter ↗openai/gpt-5.4-mini
- Official model documentation ↗Family-level model document · openai · first party
- Release source ↗direct research preview identity date
Safety evals
Rank and direction are specific to the Safety portfolio. Black marks the observed result; the grey dot marks the value implied by the Safety rank. Values stay on each source’s native scale.
| Eval / sub-eval | Rank | Value | Distribution | Source |
|---|---|---|---|---|
| AA-Omnisciencehallucination_rate | #283 / 358 | ↓0.8985 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5child_sexual_exploitation | #7 / 31 | ↓19.1 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5defamation | #3 / 31 | ↓26.1 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5hate | #4 / 31 | ↓28.9 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5indiscriminate_weapons_cbrne | #10 / 31 | ↓33.3 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5intellectual_property | #1 / 31 | ↓14.9 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5non_violent_crimes | #3 / 31 | ↓23.9 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5privacy | #2 / 31 | ↓23.4 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5sex_related_crimes | #4 / 31 | ↓19.6 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5sexual_content | #6 / 31 | ↓27.1 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5specialized_advice | #3 / 31 | ↓22.48 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5suicide_self_harm | #3 / 31 | ↓8.5 | Source ↗official | |
| Adversarial Humanities Benchmark (AHB) — Table 5violent_crimes | #3 / 31 | ↓23.9 | Source ↗official | |
| Arena Factuality — Text Arena (factuality-only weighting)factuality_bt_rating | #1 / 111 | ↑1497.0 | Source ↗official | |
| BullshitBench v2clear_pushback_rate | #65 / 122 | ↑0.3167 | Source ↗official | |
| CAIS Risk Indexagent_red_teaming | #27 / 49 | ↓74.3 | Source ↗official | |
| CAIS Risk Indexbioweapons_assistance | #15 / 54 | ↓60 | Source ↗official | |
| CAIS Risk Indexhle_overconfidence | #29 / 55 | ↓55.2 | Source ↗official | |
| CAIS Risk Indexmachiavelli | #44 / 51 | ↓95.6 | Source ↗official | |
| CAIS Risk Indexmask | #25 / 57 | ↓15.9 | Source ↗official | |
| CAIS Risk Indexpolitical_manipulation | #47 / 51 | ↓61.1 | Source ↗official | |
| CAIS Risk Indextextquests_harm | #13 / 54 | ↓15.4 | Source ↗official | |
| Cisco AI Defense Rolling Single-Turn Leaderboardsingle_turn_attack_success_rate | #26 / 104 | ↓9.571 | Source ↗official | |
| DelusionEvaldelusional_prevalence_pct | #1 / 16 | ↓11.2 | Source ↗official | |
| DelusionEvaldiscourages_harm_prevalence_pct | #2 / 16 | ↑61.3 | Source ↗official | |
| DelusionEvalfacilitates_harm_prevalence_pct | #1 / 16 | ↓0 | Source ↗official | |
| DelusionEvalrelationship_prevalence_pct | #1 / 16 | ↓7.1 | Source ↗official | |
| DelusionEvalsycophancy_prevalence_pct | #2 / 16 | ↓15.2 | Source ↗official | |
| DystopiaBenchbasaglia_score | #14 / 50 | ↓49.13 | Source ↗official | |
| DystopiaBenchbaudrillard_score | #12 / 50 | ↓38.87 | Source ↗official | |
| DystopiaBenchhuxley_score | #11 / 50 | ↓45.1 | Source ↗official | |
| DystopiaBenchlaguardia_score | #11 / 50 | ↓52.8 | Source ↗official | |
| DystopiaBenchorwell_score | #12 / 50 | ↓41.8 | Source ↗official | |
| DystopiaBenchpetrov_score | #9 / 50 | ↓39.6 | Source ↗official | |
| Every Model Cheats — Cybench Cheat Propensitybaseline_cheat_propensity_pct | #5 / 22 | ↓21.7 | Source ↗official | |
| Every Model Cheats — Cybench Cheat Propensitysevere_anti_cheat_cheat_propensity_pct | #17 / 22 | ↓13 | Source ↗official | |
| Every Model Cheats — Cybench Cheat Propensitystandard_anti_cheat_cheat_propensity_pct | #10 / 22 | ↓13 | Source ↗official | |
| Governance Decay under Passive Context Compactiongovernance_retention_score | #5 / 7 | ↑59 | Source ↗official | |
| JuICE Cultural-Error Span Detectionf1 | #9 / 10 | ↑0.3079 | Source ↗official | |
| MACHIAVELLIdeception_relative_random_pct | #41 / 50 | ↓95.6 | Source ↗official | |
| Pander Scoreconversational_absolute_pander_score | #18 / 26 | ↓16.43 | Source ↗official | |
| Pander Scoreinstructional_absolute_pander_score | #15 / 26 | ↓38.82 | Source ↗official | |
| RefusalBenchyouden_j | #5 / 19 | ↑0.3787 | Source ↗official | |
| SM-Benchadversarial | #79 / 92 | ↑76.83 | Source ↗official | |
| SM-Benchambiguous_interpretation | #52 / 92 | ↑84.08 | Source ↗official | |
| SM-Benchanti_hallucination | #81 / 92 | ↑80.23 | Source ↗official | |
| SM-Bencheq_boundaries | #38 / 92 | ↑67.28 | Source ↗official | |
| SM-Benchoverfit | #90 / 92 | ↑15.3 | Source ↗official | |
| Vals AI Cheating Auditterminal_bench_cheating_shortcut_evidence_rate_pct | #2 / 14 | ↓0.3745 | Source ↗official | |
| Vectara HHEM Factual Consistencyfactual_consistency_rate | #16 / 94 | ↑94.5 | Source ↗official | |
| VETO Misfired Alignmentmisfired_alignment_rate_pct | #17 / 25 | ↓9.9 | Source ↗official |
Freedom evals
Rank and direction are specific to the Freedom portfolio. Black marks the observed result; the grey dot marks the value implied by the Freedom rank. Values stay on each source’s native scale.
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 | 5.5 | |
| Honesty-humility | 5.3 | |
| Extraversion | 6.1 | |
| Agreeableness | 6.5 | |
| Conscientiousness | 7.3 |