‹ PublicAI Index
The LLM benchmark aggregator.
Claude Haiku 4.5
Anthropic
Strongest in Tool use (#6 of 81), weakest in Secure code (#194 of 270). Above par in 11 of 26 scopes. Among the models it meets almost everywhere, it finishes behind Claude Sonnet 5 and Claude Fable 5 and ahead of DeepSeek V3 and Mistral Large 3.
§ 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.
Safety55.4−6#44/3293/3
Jailbreak resistance61.8−3.2#25/2681/1
Harm refusal57.6−3.9#27/2922/2
Toxicity avoidance57.4−1.6#41/2681/1
Fairness57.1−12.9#49/2922/2
Factual grounding50.4−20.3#57/1041/1
Safe-prompt compliance47.9−12.5#61/821/1
Secure code44.9−21.3#194/2701/1
Agents53.2−13.2#77/2983/6
Tool use72.4−1.6#6/811/1
Knowledge work45.3−25.4#122/2082/2
Reasoning48.9−16.5#107/1952/5
Mathematics54.4−9.4#62/1521/2
Science47.3−16.2#84/1221/1
Reasoning42.7−26.3#102/1471/3
Human preference58−10.8#119/3541/1
Human preference58−10.8#119/3541/1
Professional46.3−16.1#143/1941/2
Medical47−19.5#103/1491/1
Finance46.3−16.1#116/1621/1
Legal44.5−21.4#123/1731/1
Knowledge45.8−16.7#147/1972/3
Document parsing45.5−17.8#56/831/1
Academic knowledge46.9−17#93/1231/1
Coding40.8−25.4#162/1791/6
Agentic coding40.3−26.6#158/1731/5
§ 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 Claude Haiku 4.5, left for the other.
§ 3 · Sources
Where the numbers come from
10 publications, 39 figures. Every one links to the page it was read from.
LMArena Text 1414
GDPval-AA 740
AA-Briefcase 617
ARC-AGI-2 4%
BFCL v4 68.7%
Vals · Legal Research Bench 10.58%Vals · CaseLaw 56.48%Vals · LegalBench 81.24%Vals · Finance Agent 31.01%Vals · CorpFin 60.61%Vals · TaxEval 67.54%Vals · MortgageTax 62.16%Vals · MedQA 79.57%Vals · MedCode 32.68%Vals · MedScribe 85.23%Vals · SWE-bench Verified 66.6% (Mini-SWE-agent)Vals · Vibe Code Bench 11.39% (OpenHands)Vals · Code Migration 7.55%Vals · GPQA Diamond 72.22%Vals · MMLU Pro 78.72%Vals · AIME 82.71%
ParseBench · Overall 51.21%ParseBench · Tables 78.53%ParseBench · Charts 27.78%ParseBench · Content faithfulness 83.1%ParseBench · Semantic formatting 55.53%ParseBench · Visual grounding 11.11%
HELM Safety · HarmBench 95.9%HELM Safety · SimpleSafetyTests 98.8%HELM Safety · Anthropic Red Team 96.9%HELM Safety · BBQ 92.8%HELM Safety · XSTest 93.2%
Enkrypt · Jailbreak risk 2.9%Enkrypt · Harmful content risk 1.1%Enkrypt · CBRN risk 10.5%Enkrypt · Toxicity risk 1.1%Enkrypt · Bias risk 64.9%Enkrypt · Insecure code risk 43.6%
Vectara · Factual consistency 90.2%
Badge
[](https://publicai.io/model-index/m/claude-haiku-4-5)