Skip to leaderboard

Agentic coding benchmark

Standings ·

Jungle Leaderboard

Six frontier models. Two hours each. One traditional Chinese board game.

Every model got roughly two hours of agentic coding time to build a playable Jungle (Dou Shou Qi) program from an empty repository. The finished programs were then played against one another — and any program in which a defect surfaced during competition failed.

Models evaluated
6
Build budget
~2 h each
Weights
3 open · 3 closed
Game
Dou Shou Qi

Top three

  1. First place

    DeepSeek-V4-Pro-0813

    DeepSeek · open weights

    ClaudeCode + Vision MCP

  2. Second place

    Qwen-3.8

    Alibaba · open weights

    ClaudeCode

  3. Third place

    Claude-Fable-5

    Anthropic · closed weights

    Everything-ClaudeCode

Full standings

Standard competition ranking: two models share fifth place, so no sixth place is awarded.Access indicates whether the model's weights are publicly available. Every link opens in a new tab.
RankModelAccessCoding harnessEffortArtifacts
1
DeepSeek-V4-Pro-0813
DeepSeek
OpenClaudeCode + Vision MCPofficial Max
2
Qwen-3.8
Alibaba
OpenClaudeCodexhigh
3
Claude-Fable-5
Anthropic
ClosedEverything-ClaudeCodeMax
4
Grok-4.6
xAI
ClosedGrokBuildxhigh
5tied
GPT-5.6-Sol
OpenAI
ClosedCodexUltra
5tied
Kimi-K3
Moonshot AI
OpenClaudeCodeMax

Methodology

Build phase. Every model received the same task — implement a complete, playable Jungle (Dou Shou Qi) program — and approximately 2hours of agentic coding time in which to finish it. Each model worked inside the coding harness and effort setting recorded against it in the table above, so a row describes a model-plus-harness configuration rather than a model in isolation.

Competition phase. The finished programs were then played against one another. A model is credited with a win when the program it wrote wins its match: what competes is the artifact the model produced, not the model playing live.

Failure criterion. A program in which a defect is observed during competition — an illegal move accepted, a rule applied incorrectly, a crash, or a state the game cannot recover from — is recorded as a failure for the model that wrote it. Correctness is therefore a precondition rather than a tiebreaker: a program that plays strongly but violates a rule does not outrank one that plays correctly.

Snapshot. These standings reflect results as of. Per-match records are not part of this snapshot — the table publishes final placements only. Every entry links to both the deployed build and its full source, so any result here can be replayed and inspected directly.

The game, and why it is a useful test

Jungle — 斗兽棋, Dou Shou Qi, "the game of fighting animals" — is a two-player board game played on a 7×9 grid. Each side commands eight animals. A player wins by moving any piece into the opponent's den, or by capturing every opposing piece.

A piece may ordinarily capture an enemy of equal or lower rank:

  1. Elephantrank 8
  2. Lionrank 7
  3. Tigerrank 6
  4. Leopardrank 5
  5. Wolfrank 4
  6. Dogrank 3
  7. Catrank 2
  8. Ratrank 1

Except at the very ends of that ladder: the Rat captures the Elephant, while the Elephant can never touch the Rat. The ranking is a loop, not a line.

Three more exceptions govern the board itself:

  • Only the Rat may enter the two river regions, and it cannot capture while in the water.
  • The Lion and the Tiger leap a river in a straight line — unless a rat of either colour sits anywhere along the path.
  • A piece standing in one of the opponent's traps loses its rank entirely and can be captured by anything.

These exceptions interact, which is what makes the game a searching two-hour target. A naive implementation yields something that resembles Jungle and plays legally most of the time; the defects concentrate in precisely the positions the exceptions govern. Competition play is what forces those positions to occur.

Reading the table

Places follow standard competition ranking. Two models finished level and therefore share fifth place; the place immediately below a shared position is skipped, so there is no sixth. Rows marked tied are level with one another, not ordered within the tie — the alphabetical order shown inside a tie carries no meaning.

Coding harness names the agent that drove the model, and Effort the reasoning tier it ran at. Both materially affect the result, so a row should be read as one configuration of a model rather than as a verdict on the model at every setting.

Each Play link opens that model's deployed build in a new tab, and each Source link opens the corresponding GitHub repository. All twelve destinations are third-party pages published from this benchmark's own runs.