‹ PublicAI Index
The LLM benchmark aggregator.
DeepSeek V4 Pro
DeepSeek · 1.7T · open weights
Strongest in Cybersecurity (#14 of 47), weakest in Toxicity avoidance (#211 of 268). Above par in 24 of 33 scopes. Among the models it meets almost everywhere, it finishes behind Claude Fable 5 and Claude Fable 5.1 and ahead of GPT-5 and GPT-5 mini.
§ 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.
Coding58.5−7.7#17/1792/6
Agentic coding58.9−8#15/1732/5
Code generation47.4−17.9—/20✱0/1
Repository Q&A50.5−16.2—/0✱✱
Agents56.3−10.1#43/2982/6
Knowledge work60.1−10.6#46/2082/2
Workflow automation42.9−23.1—/0✱✱
Human preference62.8−6#50/3541/1
Human preference62.8−6#50/3541/1
Reasoning54.5−10.9#54/1954/5
Science61.4−2.1#17/1221/1
Reasoning51.9−17.1#57/1472/3
Mathematics53.5−10.3#67/1522/2
Expert reasoning44.8−21.6—/0✱✱
Knowledge53−9.5#63/1971/3
Academic knowledge57.2−6.7#35/1231/1
Professional52.3−10.1#69/1941/2
Cybersecurity57.1−6.8#14/471/1
IT operations40.5−33.3#26/331/1
Finance55.1−7.3#45/1621/1
Legal53.1−12.8#67/1731/1
Medical50.8−15.7#79/1491/1
Core abilities50.7−14.6#92/1993/4
Language55.3−11.3#24/621/1
Instruction following49.5−24.3#33/621/1
General intelligence49.2−20.2#97/1782/3
Safety50.4−11#166/3292/3
Factual grounding53.4−17.3#43/1041/1
Jailbreak resistance54.4−10.6#122/2681/1
Harm refusal51.4−10.1#130/2921/2
Secure code49.7−16.5#156/2701/1
Fairness45.2−24.8#187/2921/2
Toxicity avoidance48.1−10.9#211/2681/1
§ 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 DeepSeek V4 Pro, left for the other.
§ 3 · Sources
Where the numbers come from
12 publications, 53 figures. Every one links to the page it was read from.
LMArena Text 1464
Artificial Analysis Intelligence Index 23
GDPval-AA 1457
AA-Briefcase 1253
ARC-AGI-2 61.3%
LiveBench 1.0 77.4LiveBench 1.0 · Reasoning 85.8LiveBench 1.0 · Coding 77.2LiveBench 1.0 · Agentic Coding 54.9LiveBench 1.0 · Mathematics 95.1LiveBench 1.0 · Data Analysis 79.2LiveBench 1.0 · Language 82.1LiveBench 1.0 · Instruction Following 67.7
Kagi LLM Benchmark 53.5%
SimpleBench 50.9%
Vals · Legal Research Bench 40.87%Vals · LegalBench 82.36%Vals · Harvey Legal Agent Benchmark 10.83%Vals · Finance Agent 50.39%Vals · CorpFin 65.42%Vals · TaxEval 73.06%Vals · MedCode 42.47%Vals · MedScribe 80.17%Vals · CyberBench 72.86%Vals · SRE Bench 1.91%Vals · SWE-bench Verified 96.4% (Mini-SWE-agent)Vals · Vibe Code Bench 82.3% (OpenHands)Vals · Code Migration 41.54%Vals · GPQA Diamond 92.42%Vals · MMLU Pro 86.97%Vals · ProofBench 50%
Enkrypt · Jailbreak risk 9.1%Enkrypt · Harmful content risk 22.2%Enkrypt · CBRN risk 17.3%Enkrypt · Toxicity risk 7.5%Enkrypt · Bias risk 85.3%Enkrypt · Insecure code risk 33.8%
Vectara · Factual consistency 91.4%
GPQA Diamond (Pass@1) 92.4%Codeforces (Rating) 3348MathArena Apex (Pass@1) 65.3%Terminal-Bench 2.1 (Pass@1) 87.9%Terminal-Bench 3.0 (Pass@1) 11.8%Terminal-Bench 4.0 (Pass@1) 12.4%DeepSWE v1.1 (Resolved) 62.7%ProgramBench (Almost@1) 15.5%NL2Repo-Bench (Score) 61.5%CyberGym (Pass@1) 83.3%SEC-Bench Pro (Pass@1) 56.4%ExploitGym (Pass@1) 5.4%HLE w/ tools (Pass@1) 60%AutomationBench (Pass@1) 43.2%Agent's Last Exam (Pass@1) 25.7%
Badge
[](https://publicai.io/model-index/m/deepseek-v4-pro)