‹ PublicAI Index
The LLM benchmark aggregator.
GPT-5 mini
OpenAI
Strongest in Harm refusal (#10 of 292), weakest in Toxicity avoidance (#176 of 268). Above par in 20 of 28 scopes. Among the models it meets almost everywhere, it finishes behind Claude Fable 5 and Claude Fable 5.1 and ahead of O3 and Kimi K2.6.
§ 1 · Profile
What it is good at
Bars run from 50 — the average of the models each source lists — so right of the line is above par. The middle column is the gap to whoever leads that scope.
Safety58.4−3#15/3293/3
Harm refusal59.8−1.7#10/2922/2
Secure code64.7−1.5#14/2701/1
Safe-prompt compliance57.9−2.5#15/821/1
Jailbreak resistance62.2−2.8#22/2681/1
Fairness61.2−8.8#34/2922/2
Factual grounding42.5−28.2#86/1041/1
Toxicity avoidance52−7#176/2681/1
Professional55.2−7.2#29/1941/2
Legal58.4−7.5#20/1731/1
Finance55.1−7.3#48/1621/1
Medical54.1−12.4#53/1491/1
Core abilities56−9.3#34/1991/4
General intelligence58.6−10.8#31/1781/3
Reasoning51.5−13.9#88/1952/5
Mathematics56.7−7.1#37/1521/2
Science52.9−10.6#67/1221/1
Reasoning42.8−26.2#101/1471/3
Agents49.9−16.5#128/2983/6
Tool use62.8−11.2#14/811/1
Knowledge work43.3−27.4#133/2082/2
Human preference55.7−13.1#144/3541/1
Human preference55.7−13.1#144/3541/1
Knowledge46.1−16.4#146/1972/3
Document parsing45.1−18.2#58/831/1
Academic knowledge51.3−12.6#74/1231/1
Coding42.1−24.1#153/1791/6
Agentic coding41.6−25.3#149/1731/5
§ 2 · Head to head
What it beats, and what beats it
The same models turn up scope after scope. Counted once: where both were placed, who finished higher. Bars run right for GPT-5 mini, left for the other.
§ 3 · Sources
Where the numbers come from
11 publications, 37 figures. Every one links to the page it was read from.
LMArena Text 1390
GDPval-AA 773
AA-Briefcase 429
ARC-AGI-2 4.4%
BFCL v4 55.46%
Kagi LLM Benchmark 70.3%
Vals · CaseLaw 68.49%Vals · LegalBench 81.77%Vals · CorpFin 60.18%Vals · TaxEval 75.22%Vals · MortgageTax 66.89%Vals · MedQA 96.06%Vals · MedCode 43.05%Vals · MedScribe 80.58%Vals · SWE-bench Verified 60.8% (Mini-SWE-agent)Vals · Vibe Code Bench 14.17% (OpenHands)Vals · GPQA Diamond 80.3%Vals · MMLU Pro 82.23%Vals · AIME 91.46%
ParseBench · Overall 49.86%ParseBench · Tables 74.53%ParseBench · Charts 38.9%ParseBench · Content faithfulness 86.12%ParseBench · Semantic formatting 38.05%ParseBench · Visual grounding 11.69%
HELM Safety · HarmBench 97.1%HELM Safety · SimpleSafetyTests 100%HELM Safety · Anthropic Red Team 99.1%HELM Safety · BBQ 96.3%HELM Safety · XSTest 97.7%
Enkrypt · Jailbreak risk 2.6%Enkrypt · Harmful content risk 8.9%Enkrypt · CBRN risk 3.2%Enkrypt · Toxicity risk 4.8%Enkrypt · Bias risk 58.9%Enkrypt · Insecure code risk 3.1%
Vectara · Factual consistency 87.1%
Badge
[](https://publicai.io/model-index/m/gpt-5-mini)