‹ PublicAI Index
The LLM benchmark aggregator.
GPT-5 nano
OpenAI
Strongest in Harm refusal (#4 of 292), weakest in Human preference (#189 of 354). Above par in 14 of 24 scopes. Among the models it meets almost everywhere, it finishes behind Claude Fable 5 and Claude Fable 5.1 and ahead of GPT-6 Luna and MiMo V2.6 Flash.
§ 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.
Safety59.6−1.8#9/3293/3
Harm refusal60.6−0.9#4/2922/2
Secure code65.6−0.6#7/2701/1
Jailbreak resistance63−2#10/2681/1
Safe-prompt compliance57.7−2.7#17/821/1
Fairness61.1−8.9#36/2922/2
Factual grounding48.6−22.1#62/1041/1
Toxicity avoidance54.5−4.5#119/2681/1
Core abilities53.2−12.1#65/1991/4
General intelligence54.7−14.7#53/1781/3
Agents54.1−12.3#67/2981/6
Tool use59.9−14.1#20/811/1
Reasoning46.8−18.6#122/1952/5
Mathematics54−9.8#64/1521/2
Science41.2−22.3#96/1221/1
Reasoning42.4−26.6#106/1471/3
Knowledge47.3−15.2#138/1971/3
Academic knowledge43.6−20.3#99/1231/1
Professional43.2−19.2#166/1941/2
Finance46.1−16.3#118/1621/1
Medical45.2−21.3#118/1491/1
Legal35.5−30.4#164/1731/1
Human preference50.7−18.1#189/3541/1
Human preference50.7−18.1#189/3541/1
§ 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 nano, left for the other.
§ 3 · Sources
Where the numbers come from
8 publications, 26 figures. Every one links to the page it was read from.
LMArena Text 1338
ARC-AGI-2 2.6%
BFCL v4 51.45%
Kagi LLM Benchmark 62.2%
Vals · CaseLaw 52.63%Vals · LegalBench 50.13%Vals · TaxEval 67.38%Vals · MortgageTax 53.62%Vals · MedQA 93.26%Vals · MedCode 30.44%Vals · MedScribe 72.86%Vals · GPQA Diamond 63.38%Vals · MMLU Pro 76.07%Vals · AIME 81.18%
HELM Safety · HarmBench 98.4%HELM Safety · SimpleSafetyTests 100%HELM Safety · Anthropic Red Team 99.6%HELM Safety · BBQ 97.6%HELM Safety · XSTest 97.6%
Enkrypt · Jailbreak risk 1.9%Enkrypt · Harmful content risk 7.2%Enkrypt · CBRN risk 2%Enkrypt · Toxicity risk 3.1%Enkrypt · Bias risk 61%Enkrypt · Insecure code risk 1.3%
Vectara · Factual consistency 89.5%
Badge
[](https://publicai.io/model-index/m/gpt-5-nano)