‹ PublicAI Index
The LLM benchmark aggregator.
Claude Sonnet 5
Anthropic
Strongest in Safety (#1 of 329, on 1 of its 3 boards), weakest in Core abilities (#134 of 199). Above par in 30 of 35 scopes. Among the models it meets almost everywhere, it finishes behind Claude Fable 5.1 and GPT-6 Astra and ahead of Grok 4.5 and Gemini 3.5 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.
Safety61.4leads#1/3291/3
Fairness70leads#5/2921/2
Toxicity avoidance58.7−0.3#10/2681/1
Jailbreak resistance62.9−2.1#12/2681/1
Harm refusal59.5−2#13/2921/2
Secure code61.9−4.3#46/2701/1
Agents57.1−9.3#33/2982/6
Knowledge work61.4−9.3#37/2082/2
Tool use59.4−14.6—/81✱0/1
Web research56−7.9—/35✱0/2
Workflow automation62.5−3.5—/0✱✱
Reasoning56−9.4#35/1953/5
Science58.9−4.6#33/1221/1
Mathematics56.2−7.6#42/1522/2
Reasoning53.5−15.5#51/1471/3
Expert reasoning64−2.4—/0✱✱
Coding54.7−11.5#40/1793/6
Agentic coding54.9−12#41/1733/5
Code generation64.7−0.6—/20✱0/1
Professional53.7−8.7#48/1941/2
Finance58.8−3.6#8/1621/1
IT operations39.3−34.5#32/331/1
Legal55.9−10#35/1731/1
Cybersecurity47.3−16.6#35/471/1
Medical51.9−14.6#76/1491/1
Human preference62.5−6.3#51/3541/1
Human preference62.5−6.3#51/3541/1
Knowledge53−9.5#65/1972/3
Academic knowledge57.9−6#26/1231/1
Document parsing52.2−11.1#43/831/1
Factuality60.8leads—/0✱✱
Core abilities45.9−19.4#134/1991/4
Language45.9−20.7#43/621/1
Instruction following42.8−31#45/621/1
Long context60.6−4—/0✱✱
§ 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 5, left for the other.
§ 3 · Sources
Where the numbers come from
10 publications, 56 figures. Every one links to the page it was read from.
LMArena Text 1461
GDPval-AA 1470
AA-Briefcase 1356
Terminal-Bench 12.1% (Claude Code)
LiveBench 1.0 76LiveBench 1.0 · Reasoning 88.7LiveBench 1.0 · Coding 80.7LiveBench 1.0 · Agentic Coding 59.4LiveBench 1.0 · Mathematics 92.9LiveBench 1.0 · Data Analysis 71.7LiveBench 1.0 · Language 75LiveBench 1.0 · Instruction Following 63.9
SimpleBench 60.6%
Vals · Legal Research Bench 41.83%Vals · LegalBench 83.92%Vals · Finance Agent 53.91%Vals · CorpFin 66.98%Vals · TaxEval 75.63%Vals · MortgageTax 70.03%Vals · MedCode 47.54%Vals · MedScribe 76.05%Vals · CyberBench 61.91%Vals · SRE Bench 0.38%Vals · SWE-bench Verified 79.6% (Mini-SWE-agent)Vals · Vibe Code Bench 81.33% (OpenHands)Vals · Code Migration 44.39%Vals · GPQA Diamond 88.89%Vals · MMLU Pro 87.55%Vals · ProofBench 77%
ParseBench · Overall 61.16%ParseBench · Tables 86.65%ParseBench · Charts 60.53%ParseBench · Content faithfulness 86.51%ParseBench · Semantic formatting 64.93%ParseBench · Visual grounding 7.18%
Enkrypt · Jailbreak risk 2%Enkrypt · Harmful content risk 0%Enkrypt · CBRN risk 7.5%Enkrypt · Toxicity risk 0.2%Enkrypt · Bias risk 29.5%Enkrypt · Insecure code risk 8.9%
GDPVal-AA 1584tau3-Banking 37.3%Toolathlon Verified 71.6%Automation Bench Public 34.7%Apex-Agents (pass@1) 31.7%MCPMark 65.3%BrowseComp 84.7%Terminal-Bench 2.1 80.5%SciCode 53.6%Humanity's Last Exam (without tools) 41.3%GPQA Diamond 91.1%CritPt 16.9%AA-LCR 77%AA-Omniscience Accuracy 40%AA-Omniscience Non-Hallucination 61%
Badge
[](https://publicai.io/model-index/m/claude-sonnet-5)