AIEvals

Models

DeepSeek V4 Pro

The 13 August revision of DeepSeek V4 Pro; the April release remains a separate model in source leaderboards.

DeepSeekReleased 13 August 2026Context 1M tokensOutput 384k tokensOpen weights$1.32 in / $3.96 out per Mtok

Family DeepSeek V449 benchmarks with a published resultNo first place on the independent columns among the 46 models this site listsRead 17 October 2026

Where it stands

Rank compares tracked models with a score. Published shows the source’s rank and field size, when available.

BenchmarkScoreRankPublishedLeaderBehind
Vals IndexVals AI47.63%29of 3331of 41Gemini 4 Argon21.27 pts
Artificial Analysis Intelligence IndexArtificial Analysis3638of 42no list readClaude Opus 5.522 pts
Terminal-Bench 4.0Artificial Analysis14.14%28of 40no list readClaude Sonnet 5.549.50 pts
SWE-bench (Verified)Vals AI96.40%2of 232of 88Claude Opus 50.60 pts
LiveCodeBenchVals AI87.53%10of 2113of 143Claude Fable 5.12.99 pts
GPQA DiamondVals AI92.42%14of 2316of 138Gemini 3.1 Pro3.03 pts
MMLU ProVals AI86.97%17of 2436of 138Claude Fable 5.15.41 pts
Vibe Code Bench v1.1Vals AI82.30%16of 3817of 106Claude Sonnet 5.510.09 pts
Terminal-Bench Science 0.1Artificial Analysis5.71%18of 24no list readGPT-6 Astra57.62 pts
ProgramBenchVals AI0.00%19of 2923of 54Claude Opus 5.518.50 pts
τ³-BankingArtificial Analysis39.59%17of 30no list readQwen3.8 Max11.75 pts
MLCR-AAArtificial Analysis17.78%24of 37no list readClaude Sonnet 5.557.22 pts
Vibe Code Bench 1–100Vals AI17.53%11of 2011of 20Claude Opus 5.512.83 pts
EnterpriseOps-Gym-AAArtificial Analysis49.60%3of 22no list readClaude Fable 51.52 pts
AutomationBench-AAArtificial Analysis56.71%23of 40no list readGemini 4 Argon20.80 pts
SciCodeArtificial Analysis51.04%36of 40no list readClaude Opus 5.515.86 pts
GDP.pdfArtificial Analysis11.40%36of 40no list readGPT-6.1 Sol19.60 pts
CritPtArtificial Analysis18.00%27of 42no list readGPT-5.6 Sol14.29 pts
AA-OmniscienceArtificial Analysis0.8335of 42no list readClaude Opus 5.545.59
AA-LCR v1.1Artificial Analysis80.33%26of 42no list readKimi K38.34 pts
Terminal-Bench 2.1Vals AI54.68%32of 3550of 75Claude Opus 5.532.96 pts
ProofBench v1.1Vals AI50.00%24of 3428of 46Claude Sonnet 5.550.00 pts
LegalBenchVals AI82.36%27of 2771of 149Claude Fable 56.20 pts
Harvey's Legal Agent BenchmarkVals AI7.50%17of 3722of 73Muse Spark 1.217.92 pts
SkillsBenchVals AI53.83%13of 2016of 35DeepSeek V4.1 Flash15.97 pts
Code MigrationVals AI41.54%25of 3725of 71Claude Sonnet 5.528.29 pts
CyberBench v1.1Vals AI72.86%9of 3111of 43GPT-6 Sol5.12 pts
EMBVals AI52.80%33of 3746of 69Claude Fable 5.123.87 pts
Finance Agent (v2)Vals AI50.39%30of 3737of 73Gemini 4 Argon15.01 pts
IOIVals AI51.61%22of 3123of 38Gemini 4 Argon48.39 pts
Legal Research BenchVals AI40.87%22of 3723of 72Muse Spark 1.3 Max14.42 pts
MedCodeVals AI42.47%29of 3548of 104Claude Opus 521.10 pts
MedScribeVals AI80.17%33of 3656of 106Claude Opus 5.511.26 pts
Tax Agent BenchVals AI58.66%30of 3735of 64Claude Fable 5.118.98 pts
DeepSWE v1.1Datacurve62.83%14of 2014of 28GPT-6 Astra11.29 pts
GDPval-AA v2.1Artificial Analysis1,441.4229of 42no list readClaude Opus 5.5404.75
AA-Briefcase v1.1Artificial Analysis1,257.6933of 41no list readClaude Opus 5.5564.16
Humanity's Last ExamArtificial Analysis41.01%33of 42no list readClaude Opus 5.520.34 pts

Nearby on Vals Index

The two models above and below on the Vals Index (Vals AI), on every column where this model has a published number.

ModelVals Index%Artificial Analysis Intelligence IndexindexTerminal-Bench 4.0%SWE-bench%LiveCodeBench%GPQA Diamond%MMLU Pro%Vibe Code Bench v1.1%Terminal-Bench Science 0.1%ProgramBench%
Muse Spark 1.2Meta49.29%407.07%86.60%––88.28%79.10%––
Qwen3.8 MaxAlibaba48.27%4018.69%85.60%87.85%93.69%88.60%64.70%–0.00%
DeepSeek V4 ProDeepSeek47.63%3614.14%96.40%87.53%92.42%86.97%82.30%5.71%0.00%
Grok 4.5xAI44.72%3910.61%86.60%87.35%92.93%89.22%69.00%–0.00%
MiniMax M3MiniMax36.53%292.02%75.00%82.15%92.68%84.22%47.57%0.48%–
Bold marks a figure better than this model’s on the same column, whichever model it belongs to. Ranks cover the models shown from the stated source.

What the hosted runtime reports

Ollama Cloud serves its own copy of this model, so its context cap and its prices are the ones a caller actually gets. Read 25 September 2026.

Parameters
1.65T parameters49B active per token
Hosted context
1M tokens
Hosted price / Mtok
$1.32in · $0.044 cached · $3.96 out · off-peak $0.66 in, $1.98 out
Tool callingReasoning modeServed by the hostInput: Text

Three reasoning modes are exposed by the host — no thinking, thinking and max thinking — and its price halves outside peak hours.

ollama.com/library/deepseek-v4-pro

Results by benchmark

Independent and vendor results retain their own harnesses and effort settings.

BenchmarkIndependent runAs reported by a publisher
τ³-BankingTool use39.59%Artificial Analysisnot published
MLCR-AAMedical reasoning17.78%Artificial Analysisnot published
Vibe Code Bench 1–100Application development17.53%Vals AInot published
EnterpriseOps-Gym-AAEnterprise workflows49.60%Artificial Analysisnot published
AutomationBench-AAWorkflow automation56.71%Artificial Analysisnot published
SciCodeScientific coding51.04%Artificial Analysisnot published
GDP.pdfDocument reasoning11.40%Artificial Analysisnot published
CritPtScientific reasoning18.00%Artificial Analysisnot published
AA-OmniscienceFactual knowledge0.83Artificial Analysisnot published
AA-LCR v1.1Long-context reasoning80.33%Artificial Analysisnot published
Terminal-Bench 4.0Agentic coding14.14%Artificial Analysis
14.14%Vals AI
12.4%Terminal-Bench 4.0 (Pass@1)DeepSeek · 10 September 2026Rank 6 of 7 in this publisher table
Terminal-Bench 2.1Agentic coding54.68%Vals AI
78.65%Artificial Analysis
87.9%Terminal Bench 2.1Z.ai · 14 August 2026Rank 5 of 8 in this publisher table
87.9%Terminal-Bench 2.1 (Pass@1)DeepSeek · 10 September 2026Rank 6 of 7 in this publisher table
SWE-bench (Verified)Software engineering96.40%Vals AInot published
LiveCodeBenchCode generation87.53%Vals AInot published
GPQA DiamondScience reasoning92.42%Vals AI
92.83%Artificial Analysis
92.4%GPQA Diamond (Pass@1)DeepSeek · 10 September 2026Rank 4 of 7 in this publisher table
MMLU ProKnowledge86.97%Vals AInot published
Vibe Code Bench v1.1Product coding82.30%Vals AInot published
Terminal-Bench Science 0.1Agentic science5.71%Artificial Analysis
0.00%Vals AI
not published
ProgramBenchAgentic coding0.00%Vals AI
15.5%ProgramBench (Almost@1)DeepSeek · 10 September 2026Rank 6 of 6 in this publisher table
ProofBench v1.1Mathematics50.00%Vals AInot published
LegalBenchProfessional knowledge82.36%Vals AInot published
Harvey's Legal Agent BenchmarkProfessional knowledge7.50%Vals AInot published
SkillsBenchAgentic tooling53.83%Vals AInot published
Vals IndexComposite47.63%Vals AInot published
Code MigrationAgentic coding41.54%Vals AInot published
CyberBench v1.1Security72.86%Vals AInot published
EMBFinancial modelling52.80%Vals AInot published
Finance Agent (v2)Financial analysis50.39%Vals AInot published
IOICompetitive programming51.61%Vals AInot published
Legal Research BenchProfessional knowledge40.87%Vals AInot published
MedCodeHealthcare administration42.47%Vals AInot published
MedScribeHealthcare administration80.17%Vals AInot published
Tax Agent BenchProfessional knowledge58.66%Vals AInot published
AutomationBenchBusiness workflowsno published result
43.2%AutomationBench (v1.0.6)Z.ai · 14 August 2026Rank 5 of 8 in this publisher table
43.2%AutomationBench (Pass@1)DeepSeek · 10 September 2026Rank 6 of 7 in this publisher table
DeepSWE v1.1Software engineering62.83%Datacurve
62.7%DeepSWE (v1.1)Z.ai · 14 August 2026Rank 5 of 8 in this publisher table
62.7%DeepSWE v1.1 (Resolved)DeepSeek · 10 September 2026Rank 6 of 7 in this publisher table
GDPval-AA v2.1Knowledge work1,441.42Artificial Analysisnot published
AA-Briefcase v1.1Knowledge work1,257.69Artificial Analysisnot published
Humanity's Last ExamMultidisciplinary reasoning41.01%Artificial Analysis
60.0%HLE w/ ToolsZ.ai · 14 August 2026Rank 4 of 8 in this publisher table
42.7%HLE (Pass@1)DeepSeek · 10 September 2026Rank 4 of 7 in this publisher table
60.0%HLE w/ tools (Pass@1)DeepSeek · 10 September 2026Rank 4 of 6 in this publisher table
Agents' Last ExamComputer useno published result
25.7%Agents' Last Exam (ALE-CLI)Z.ai · 14 August 2026Rank 5 of 8 in this publisher table
25.7%Agents' Last Exam (Pass@1)DeepSeek · 10 September 2026Rank 6 of 7 in this publisher table
Artificial Analysis Intelligence IndexComposite36Artificial Analysisnot published
Terminal-Bench 3.0Agentic codingno published result
11.8%Terminal-Bench 3.0 (Pass@1)DeepSeek · 10 September 2026Rank 6 of 7 in this publisher table
NL2Repo-BenchAgentic codingno published result
61.1%NL2RepoZ.ai · 14 August 2026Rank 2 of 6 in this publisher table
61.5%NL2Repo-Bench (Score)DeepSeek · 10 September 2026Rank 3 of 7 in this publisher table
CyberGymSecurityno published result
83.3%CyberGymZ.ai · 14 August 2026Rank 4 of 8 in this publisher table
83.3%CyberGym (Pass@1)DeepSeek · 10 September 2026Rank 4 of 6 in this publisher table
SEC-bench ProSecurityno published result
56.4%SEC-Bench Pro (Pass@1)DeepSeek · 10 September 2026Rank 3 of 4 in this publisher table
ExploitGymSecurityno published result
5.4%ExploitGym (Pass@1)DeepSeek · 10 September 2026Rank 5 of 6 in this publisher table
MathArena ApexMathematicsno published result
65.3%MathArena Apex (Pass@1)DeepSeek · 10 September 2026Rank 3 of 4 in this publisher table
Codeforces (rating)Competitive programmingno published result
3,348Codeforces (Rating)DeepSeek · 10 September 2026Rank 2 of 3 in this publisher table
Toolathlon VerifiedAgentic toolingno published result
74.1%Toolathlon VerifiedZ.ai · 14 August 2026Rank 5 of 8 in this publisher table
GDPval-AA v2Knowledge workno published result
1,590GDPval-AA v2Z.ai · 14 August 2026Rank 6 of 8 in this publisher table
Independent column: the suite named under each figure, one harness per suite. Publisher column: each lab’s own table, reproduced on the benchmark’s page. Publisher ranks compare reported scores in the same table row only; ties share a rank and unreported cells are excluded.

Sources

Read on 17 October 2026.