DeepSeek V4 Pro: benchmark results
DeepSeek's open 1600B V4 Pro model for high-end reasoning and general tasks. Provider: DeepSeek. Released 2026-04-24. Access: Open.
Aggregate Rating 1761 ± 2, rank #102 of 1656 rated models, from 2323 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AdvancedMathBench - VerifierBench | 65.1 | Meta-verification balanced F1 (%): a validity verdict counts | 100 |
| AfriEconQA | 54.5 | Token F1 (%; 862 held-out test questions, answered from the | 100 |
| AfriEconQA (Closed-Book) | 26.7 | Token F1 (%; 862 held-out test questions, question only with | 100 |
| AfriEconQA - Comparison | 50.8 | Token F1 (%; the comparison questions (across periods, count | 100 |
| AfriEconQA - List | 10.1 | Set F1 (%; the list questions (every factor, risk or interve | 100 |
| AfriEconQA - Multiple Choice | 88.4 | F1 (%; the four-option multiple-choice questions (answered w | 100 |
| AfriEconQA - Synthesis | 43.3 | Token F1 (%; the synthesis questions (connected policy claim | 100 |
| AgentS4D - Task Completion (Codex CLI) | 96.65 | Task completion rate (%): runs whose case-specific completio | 100 |
| AgentS4D - Task Completion (Hermes Agent) | 98.48 | Task completion rate (%): runs whose case-specific completio | 100 |
| AgentS4D - Task Completion (OpenClaw) | 97.26 | Task completion rate (%): runs whose case-specific completio | 100 |
| AppWorld-UL (Function Calling) - Infeasible Tasks | 33.3 | Interactive task goal completion (%; share of 102 tasks made | 100 |
| AuthMem-Bench (Unauthorized Action Rate) | 38.6 | Attack success rate (%; share of 350 non-authorizing cases i | 100 |
Interactive version: theaggregate.ai/model?slug=deepseek-v4-pro · How It Works · Data refreshed daily, snapshot 2026-10-20.