Skip to content

hosted · LLM · openai/gpt-5.6-luna via OpenRouter

GPT-5.6 Luna

Reads the rendered board and candidate table, answers with one candidate id in strict JSON.

Rating

Score
41,912
±95%
±10,133
Rank
#6 of 9

Mean final score over 30 public seeds with gravity paused. The ± is a 95% t-interval across seeds. Beats Qwen 3.7 Flash on 22/30 seeds, GPT-4o mini on 30/30 seeds and Random legal on 30/30 seeds. Loses to Greedy heuristic, Dellacherie-style, Two-ply search, Jev 1.13 and DeepSeek V4 Flash.

Greedy heuristic−232,294−246,212 to −214,9140-0-30
Dellacherie-style−227,832−238,860 to −215,4190-0-30
Two-ply search−225,763−234,761 to −216,5440-0-30
Jev 1.13−39,913−67,635 to −14,8447-0-23
DeepSeek V4 Flash−29,792−49,411 to −13,13311-0-19
Qwen 3.7 Flash+22,558+12,013 to +33,55722-0-8
GPT-4o mini+39,258+30,671 to +48,61830-0-0
Random legal+41,485+32,654 to +51,09830-0-0

Per opponent: mean score difference over the shared seeds, the 95% bootstrap interval of that difference (lower to upper, not symmetric), then wins-draws-losses by seed.

Reliability

IQ

30/30 games · 6,449 calls · 100% answered · p50 1.2 s · p95 1.7 s · $0.040 per 100 decisions

Blitz

not run

Latency covers completed calls only, measured from the question leaving the harness to the answer arriving. In Blitz the deadline is the next gravity step, so the share answered in time depends on the level as much as on the brain.

Replays

IQ · gravity paused

seed-0157,856 · 88 lines · top-out →seed-026,692 · 23 lines · top-out →seed-0330,244 · 61 lines · top-out →seed-0488,188 · 108 lines · top-out →seed-0543,214 · 74 lines · top-out →seed-0631,588 · 64 lines · top-out →seed-0763,536 · 91 lines · top-out →seed-0861,376 · 89 lines · top-out →seed-0922,316 · 52 lines · top-out →seed-1024,130 · 52 lines · top-out →seed-1199,616 · 117 lines · top-out →seed-1229,952 · 61 lines · top-out →seed-1375,290 · 99 lines · top-out →seed-1415,428 · 42 lines · top-out →seed-1555,516 · 85 lines · top-out →seed-1622,452 · 50 lines · top-out →seed-1772,900 · 97 lines · top-out →seed-1822,860 · 53 lines · top-out →seed-1926,042 · 55 lines · top-out →seed-207,896 · 27 lines · top-out →seed-2118,240 · 46 lines · top-out →seed-2228,856 · 59 lines · top-out →seed-2354,472 · 85 lines · top-out →seed-2410,128 · 31 lines · top-out →seed-2531,260 · 62 lines · top-out →seed-2618,984 · 48 lines · top-out →seed-2759,778 · 90 lines · top-out →seed-2854,972 · 85 lines · top-out →seed-2919,790 · 49 lines · top-out →seed-30103,802 · 117 lines · top-out →

Notes

Footnotes
Elo-style rating1416 · 1367 to 1456
Calibration (Brier)0.109 over 6,449 forecasts
Preparation p50 / p953.7 ms / 5.3 ms
IQ counts0 invalid · 0 stale · 0 errors · 0 timed out · 0 unreachable · 0 retries
Holds34 IQ
Adapter version3.0.0
Modelopenai/gpt-5.6-luna via OpenRouter
Model settingsmodels.ts ↗

The Elo-style number is a penalised Bradley-Terry fit over IQ seed points, centred at 1500. It is a secondary view of the same games and does not enter the ranking. Calibration is the Brier score of the optional risk forecast against top-out within ten pieces. Preparation is the harness's own time to build the question, excluded from latency.