ACBench v0.1

14 Nov 2026 · Community

Win rate vs built-in bots, 0–100 · error bars: 95% range · all via OpenRouter† DeepSeek V4.1 Flash: authored on a later game build.
  • Claude Opus 5.5 leads at 18, not clearly ahead of GPT-6.1 Sol (13).
  • All models score below the built-in starters (37).

Win rate by opponent

Win rate 0–100 against each opponent, weakest first. The gap to 100 at the hardest tier is the headroom. No bar means unscored, not zero.

Dashed line: the built-in starters against that opponent.† DeepSeek V4.1 Flash: authored on a later game build.

Run-to-run variance

Each model writes its policy several times from scratch; the spread shows how repeatable its score is.

Dot: one authoring run · line: lowest to highest · tick: mean · band: ±1σ. Unscored runs are left out, not zero.† DeepSeek V4.1 Flash: authored on a later game build.

Results and evidence

Clear
Systems selected
2
One exact comparison cohort
Independent runs
4 / 4
0 failed authoring runs retained
Scored games
120 / 120
Scored / expected · not independent runs
Priced systems
2
Complete authoring USD · unknown is not zero
Fixed-reference scores · 0–100 · descriptive ordering, not a significance test
Independent runsScored / expected gamesEvidence
Claude Opus 5.5anthropic/claude-opus-5.5 · via OpenRouter18.316.7–20.0
Complete
2/20 failed · 3 blocks60/60$8.18Evaluation compute excludedcommunity · runner-observedRun evidence
GPT-6.1 Solopenai/gpt-6.1-sol · via OpenRouter13.30.0–26.7
Complete
2/20 failed · 3 blocks60/60$2.39Evaluation compute excludedcommunity · runner-observedRun evidence
Comparison rules & frozen release identity

Official comparisons hold harness/version, prompt/tool profile and budget fixed. Alternate harnesses use separate system-comparison suites. Filters retain the exact suite and configuration. Neither patches nor budgets are pooled.

Optional catalogue suggestions are discovery metadata, not authorship evidence.

community · runner-observed · revision 1. Published 2026-11-14T23:38:19.698Z · public release 2026-11-14T23:38:19.698Z.

SHA-256 f90e07425383f91c3bcb58e64fe49aa745555f60c9b96c7eb2599ebf4a2b1a1e

Cite: ACBench, ACBench v0.1, acbench-v0-1, revision 1, 2026-11-14T23:38:19.698Z.