‹ PublicAI Index
The LLM benchmark aggregator.
Claude Sonnet 4.5 (High, 32K, 20250929)
Anthropic
Strongest in Tool use (#2 of 81), weakest in Secure code (#205 of 273). Above par in 22 of 28 scopes. Among the models it meets almost everywhere, it finishes behind Claude Opus 4.8 and Claude Fable 5 and ahead of Grok 4 and Qwen3.7 Max.
§ 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.
Safety56.5−4.9#32/3313/3
Fairness65−5#24/2962/2
Harm refusal57.6−3.9#26/2962/2
Safe-prompt compliance55.1−5.3#34/821/1
Jailbreak resistance60.1−4.9#43/2711/1
Toxicity avoidance57.1−1.9#46/2711/1
Factual grounding44.9−25.9#76/1021/1
Secure code42.6−23.7#205/2731/1
Agents56.4−9.5#36/2894/6
Tool use74leads#2/811/1
Computer use57.6−10.1#24/811/2
Knowledge work48.9−22.9#101/1982/2
Professional54.4−8.7#36/1881/2
Medical55.5−11#38/1491/1
Legal55.1−10.8#43/1641/1
Finance54.1−8.4#63/1611/1
Knowledge53.2−9.5#52/1911/3
Academic knowledge57.7−6.2#28/1231/1
Human preference62.1−6.5#56/3501/1
Human preference62.1−6.5#56/3501/1
Reasoning52.1−12.8#78/1903/5
Mathematics55.8−8.3#43/1501/2
Science53.9−9.6#63/1221/1
Reasoning48.3−18.4#77/1442/3
Core abilities51.8−13.4#79/2251/4
General intelligence52.6−16.2#78/2141/3
Coding45.7−20.7#119/1761/6
Agentic coding45.5−21.5#114/1691/5
§ 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 Claude Sonnet 4.5 (High, 32K, 20250929), left for the other.
§ 3 · Sources
Where the numbers come from
12 publications, 33 figures. Every one links to the page it was read from.
LMArena Text 1456
GDPval-AA 910
AA-Briefcase 713
ARC-AGI-2 13.6%
BFCL v4 73.24%
OSWorld 62.9%
Kagi LLM Benchmark 57.9%
SimpleBench 54.3%
Vals · CaseLaw 62.16%Vals · LegalBench 84.08%Vals · CorpFin 61.97%Vals · TaxEval 73.3%Vals · MortgageTax 63.99%Vals · MedQA 94.71%Vals · MedCode 44.13%Vals · MedScribe 84.1%Vals · SWE-bench Verified 70% (Mini-SWE-agent)Vals · Vibe Code Bench 22.62% (OpenHands)Vals · GPQA Diamond 81.63%Vals · MMLU Pro 87.36%Vals · AIME 88.19%
HELM Safety · HarmBench 92%HELM Safety · SimpleSafetyTests 100%HELM Safety · Anthropic Red Team 98.2%HELM Safety · BBQ 98.9%HELM Safety · XSTest 96.4%
Enkrypt · Jailbreak risk 4.3%Enkrypt · Harmful content risk 0.6%Enkrypt · CBRN risk 13%Enkrypt · Toxicity risk 1.3%Enkrypt · Bias risk 52.5%Enkrypt · Insecure code risk 48.4%
Vectara · Factual consistency 88%
Badge
[](https://publicai.io/model-index/m/claude-sonnet-4-5-high-32k-20250929)