Model profile
Gemini 3.5 Flash
Evidence summary
Gemini 3.5 Flash has an estimated overall rank of #45; its 90% source-sensitivity interval is #26–#191. Its behavior-only rank is #55; company governance moves the combined estimate to #45. Published evidence spans 23 evals and 7 of 7 behavior components. Its strongest relative result is SM-Bench (overfit, #1 of 79); its weakest is SpeciEval (speciesism, #112 of 113).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗gemini-3-5-flash
- OpenRouter ↗google/gemini-3.5-flash
- Official model documentation ↗Family-level model document · google · first party
- Release source ↗curated outward evidence
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 | Source |
|---|---|---|---|---|
| AA-Omnisciencehallucination_rate | #80 / 328 | ↓0.6176 | Source ↗official | |
| ANIMAscore | #11 / 22 | ↑0.7004 | Source ↗official | |
| Arena Factuality — Text Arena (factuality-only weighting)factuality_bt_rating | #51 / 112 | ↑1440.0 | Source ↗official | |
| BioSecBench-Refusalbalanced_refusal_score | #1 / 10 | ↑0.5041 | Source ↗official | |
| BullshitBench v2clear_pushback_rate | #66 / 106 | ↑0.195 | Source ↗official | |
| CAIS Risk Indexagent_red_teaming | #21 / 44 | ↓72.3 | Source ↗official | |
| CAIS Risk Indexbioweapons_assistance | #32 / 49 | ↓78.3 | Source ↗official | |
| CAIS Risk Indexhle_overconfidence | #22 / 50 | ↓53.5 | Source ↗official | |
| CAIS Risk Indexmachiavelli | #31 / 46 | ↓92.3 | Source ↗official | |
| CAIS Risk Indexmask | #47 / 52 | ↓57.4 | Source ↗official | |
| CAIS Risk Indexpolitical_manipulation | #26 / 33 | ↓53.4 | Source ↗official | |
| CAIS Risk Indextextquests_harm | #33 / 49 | ↓20.7 | Source ↗official | |
| DystopiaBenchbasaglia_score | #45 / 50 | ↓70.93 | Source ↗official | |
| DystopiaBenchbaudrillard_score | #39 / 50 | ↓66.63 | Source ↗official | |
| DystopiaBenchhuxley_score | #43 / 50 | ↓79.23 | Source ↗official | |
| DystopiaBenchlaguardia_score | #44 / 50 | ↓70.9 | Source ↗official | |
| DystopiaBenchorwell_score | #45 / 50 | ↓75.17 | Source ↗official | |
| DystopiaBenchpetrov_score | #26 / 50 | ↓72.97 | Source ↗official | |
| Governance Decay under Passive Context Compactiongovernance_retention_score | #2 / 7 | ↑96 | Source ↗official | |
| Gray Swan indirect prompt injection (15 attempts)attack_success_probability_k15_pct | #12 / 13 | ↓60.5 | Source ↗official | |
| HUMAINE Trust, Ethics and Safetytrust_ethics_safety_score | #2 / 54 | ↑29.37 | Source ↗official | |
| MACHIAVELLIdeception_relative_random_pct | #32 / 50 | ↓92.3 | Source ↗official | |
| MORUscore | #5 / 13 | ↑77.89 | Source ↗official | |
| Olam Social Poker — Social Lie Ratesocial_lie_rate_per_10000_turns | #8 / 19 | ↓4 | Source ↗official | |
| PHAREhallucination_resistance_diagnostic | #12 / 70 | ↑0.8218 | Source ↗official | |
| PHAREharm_resistance_diagnostic | #19 / 70 | ↑0.9594 | Source ↗official | |
| PHAREjailbreak_resistance_diagnostic | #46 / 67 | ↑0.4264 | Source ↗official | |
| SM-Benchadversarial | #1 / 79 | ↑92.2 | Source ↗official | |
| SM-Benchambiguous_interpretation | #37 / 79 | ↑86.01 | Source ↗official | |
| SM-Benchanti_hallucination | #51 / 79 | ↑88.48 | Source ↗official | |
| SM-Bencheq_boundaries | #25 / 79 | ↑68.54 | Source ↗official | |
| SM-Benchoverfit | #1 / 79 | ↑98.36 | Source ↗official | |
| SpeciEvalbelief_animal_sentience | #72 / 113 | ↑6.68 | Source ↗official | |
| SpeciEvalland_animal_4ns | #3 / 113 | ↓3.62 | Source ↗official | |
| SpeciEvalsea_animal_4ns | #36 / 113 | ↓4.65 | Source ↗official | |
| SpeciEvalspeciesism | #112 / 113 | ↓3.88 | Source ↗official | |
| TACbase_welfare_rate | #69 / 76 | ↑17.31 | Source ↗official | |
| ToolPrivacyBenchprivate_mt_poi | #1 / 9 | ↓19.19 | Source ↗official | |
| ToolPrivacyBenchpublic_mt_poi | #8 / 9 | ↓19.86 | 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.