RLTasksby ulam.ai
Models / DeepSeek

DeepSeek / MODEL EVIDENCE

DeepSeek V4 Pro

28 benchmarks · 42 observations · 28 sources · released 2026-04-24

Scores at a glance

28 benchmarks · large dot: DeepSeek V4 Pro; small dots: closest models where reported

GPQA Diamond ↗90.91%
90.9190.91

DeepSeek v4 Pro (high) · GPQA diamond · Best score (across scorers) · % ↗

2 more reported results

89.65% · DeepSeek v4 Pro (max) · GPQA diamond · Best score (across scorers) · % · Epoch AI · GPQA diamond ↗

73.23% · DeepSeek v4 Pro (unknown thinking) · GPQA diamond · Best score (across scorers) · % · Epoch AI · GPQA diamond ↗

CritPt ↗12.86%
12.8612.86

DeepSeek v4 Pro (max) · CritPt · Accuracy · % ↗

1 more reported result

10% · DeepSeek v4 Pro (high) · CritPt · Accuracy · % · Epoch AI · CritPt ↗

LMCA ↗41.2%
2.7868.26

DeepSeek v4 Pro (max) · LMCA · Score · % ↗

DTBench ↗90.67%
41.6598.93

DeepSeek v4 Pro (max) · DTBench · Accuracy · % ↗

ForecastBench ↗60.9 Brier index
46.467

deepseek-v4-pro† · Overall difficulty-adjusted Brier index ↗

Closest models by score

Similar positions on shared benchmark leaderboards. Smaller gaps are closer.

31.9 percentile-point gap

8 of 11 shared benchmark comparisons

36.6 percentile-point gap

8 of 11 shared benchmark comparisons

How similarity is measured

Mean absolute gap in percentile positions on shared benchmark settings; lower is closer. At least 60% overlap is required, with a minimum of three when available. Independent evaluations are used when there are at least three; otherwise published report comparisons are used. Repeated model observations within one setting are averaged before ranking. Writing-style traits and composite indices are excluded. This describes similarity, not an overall quality ranking.

Filter comparisons to this model →

Sources

Every plotted observation links to its evaluation. Reported configurations, benchmark versions and task subsets can differ; similarity describes shared scores, not training lift.

  1. Arena leaderboard — WebDev evaluation ↗

    Arena

  2. SciCode Benchmark | Artificial Analysis ↗

    Artificial Analysis

  3. FrontierCode Leaderboard ↗

    Cognition

  4. DTBench — Conceptual Reasoning Index ↗

    Conceptual Reasoning Index

  5. LMCA — Conceptual Reasoning Index ↗

    Conceptual Reasoning Index

  6. Ale Bench — methodology and published results ↗

    Epoch AI · 2026

  7. Chess Puzzles ↗

    Epoch AI

  8. CL-bench Life ↗

    Epoch AI

  9. CritPt v1 ↗

    Epoch AI

  10. Forecastbench — methodology and published results ↗

    Epoch AI · 2026

  11. FrontierMath Tier 4 (v2) ↗

    Epoch AI

  12. FrontierMath Tiers 1-3 (v2) ↗

    Epoch AI

  13. GPQA Diamond ↗

    Epoch AI

  14. Mystery Game Puzzles ↗

    Epoch AI

  15. OTIS Mock AIME 2024-2025 ↗

    Epoch AI

  16. SimpleQA Verified ↗

    Epoch AI

  17. SWE-bench Verified ↗

    Epoch AI

  18. Vending Bench 2 — methodology and published results ↗

    Epoch AI · 2026

  19. AI benchmarks & capabilities ↗

    Epoch AI

  20. WeirdML ↗

    Håvard Tveit Ihle

  21. DistressBench — evaluation methodology and results ↗

    Scale Labs / SEAL · 2026

  22. DrugDiscoveryBench — evaluation methodology and results ↗

    Scale Labs / SEAL · 2026

  23. HiL-Bench (Human-in-Loop Benchmark) — evaluation methodology and results ↗

    Scale Labs / SEAL · 2026

  24. Scale Labs — evaluation methodology and results ↗

    Scale Labs / SEAL · 2026

  25. Scale Labs — evaluation methodology and results ↗

    Scale Labs / SEAL · 2026

  26. SWE Atlas - Test Writing — evaluation methodology and results ↗

    Scale Labs / SEAL · 2026

  27. ProofBench — evaluation methodology and results ↗

    Vals AI · 2026

  28. Surface Evolver Bench ↗

    yhenon

Names used in source tables

DeepSeek V4 Pro · DeepSeek V4 Pro (Mini-SWE-Agent) · DeepSeek V4 Pro (mini-SWE-agent) xhigh · DeepSeek v4 Pro (high) · DeepSeek v4 Pro (max) · DeepSeek v4 Pro (unknown thinking) · DeepSeek-V4-Pro · Deepseek V4 Pro · Deepseek V4 Pro (Mini-SWE-Agent) · deepseek-v4-pro · deepseek-v4-pro-high · deepseek-v4-pro-no-thinking · deepseek-v4-pro†