Acropolis

DeepSeek V4-Flash 0731

DeepSeekGrade B. Evidence grade B of A to D: good.Open weightsReleased 31 Jul 2026Domain specialist variance
Compare with peers

Domain profile

Median within-board percentile by capability domain across boards with sufficient peer overlap.

  • Agentic (2 boards): 13 (scale 0 to 100)Agentic (2 boards)13
  • Coding (2 boards): 13 (scale 0 to 100)Coding (2 boards)13
  • General text (1 boards): 68 (scale 0 to 100)General text (1 boards)68
  • Science (2 boards): 76 (scale 0 to 100)Science (2 boards)76

Benchmark appearances

Best verified run per board, with confidence intervals and Pareto frontier status.

BenchmarkRankScoreCostStatusObserved
FrontierCode

Epoch AI Benchmarking Hub · high · devin-cli

#41 / 51
18.8%UnknownStandardNot stated
GPQA diamond

Epoch AI Benchmarking Hub · max

#42 / 338
91%[87.6% to 94.5%]UnknownStandard2 Aug 2026
ARC-AGI-2

Epoch AI Benchmarking Hub · max

#96 / 299
61.4%UnknownStandardNot stated
ARC-AGI-2

Epoch AI Benchmarking Hub · high

#115 / 299
56%UnknownStandardNot stated
Bug Hunt Bench

Bug Hunt Bench · Claude Code

#115 / 145
14$1.52Standard1 Aug 2026
ARC-AGI-2

Epoch AI Benchmarking Hub · low

#131 / 299
46%UnknownStandardNot stated
ARC-AGI-2

Epoch AI Benchmarking Hub · none

#235 / 299
2.1%UnknownStandardNot stated

Reasoning effort scaling

Same benchmark and harness measured across multiple reasoning effort settings.

  • ARC-AGI-2

    Epoch AI Benchmarking Hub

    Monotone. Evidence grade A of A to D: strongest.
    none

    2.1

    low

    46.0

    high

    56.0

    max

    61.4

Closest head-to-head rivals

Direct pairwise win rates for DeepSeek V4-Flash 0731 across shared benchmark boards.