Can a model find the exact best move in three dimensions?
Models receive only the 15 × 15 × 15 board, rack, coordinate convention, and rules. One raw JSON move is scored against the exact solver optimum.
Models tested8
Completed runs14
Dataset positions6
Best point score diagnostic94.7%
Compare every run
Overview charts use point score as a diagnostic. Exact-optimal accuracy is reserved for each model’s run detail.
Performance by model
Point score · diagnostic
01
gpt-5.6-sol (high)94.7%02gpt-5.6-sol (medium)81.7%03gpt-5.5 (high)81.4%04
deepseek/deepseek-v4-pro-0813 (high)77.6%05gpt-5.6-sol (low)76.4%06gpt-5.6-luna (medium)63.4%07gpt-5.5 (medium)62.7%08deepseek/deepseek-v4-flash-0731 (high)61.7%09deepseek/deepseek-v4-pro (high)58.1%10gpt-5.6-luna (low)47.0%11gpt-5.5 (low)20.0%12
opencode/zai-coding-plan/glm-5.2 (low)14.7%13opencode/zai-coding-plan/glm-5.1 (low)0.0%14gpt-5.6-luna (high)0.0%
gpt-5.6-sol (high)94.7%02gpt-5.6-sol (medium)81.7%03gpt-5.5 (high)81.4%04deepseek/deepseek-v4-pro-0813 (high)77.6%05gpt-5.6-sol (low)76.4%06gpt-5.6-luna (medium)63.4%07gpt-5.5 (medium)62.7%08deepseek/deepseek-v4-flash-0731 (high)61.7%09deepseek/deepseek-v4-pro (high)58.1%10gpt-5.6-luna (low)47.0%11gpt-5.5 (low)20.0%12opencode/zai-coding-plan/glm-5.2 (low)14.7%13opencode/zai-coding-plan/glm-5.1 (low)0.0%14gpt-5.6-luna (high)0.0%Score vs. release time
Score vs. average tokens
Token range
Price vs. score
Model runs
Ranked by point score. This leaderboard is diagnostic; exact-optimal results remain inside each model run.
Rank / modelPoint scoreTop-20 moveLegalAvg. latencyTokens
01gpt-5.6-solhigh reasoning · full · 6 boards94.7%393 / 415 pts66.7%4 / 6100.0%6 / 6399.2sper board210,576total02gpt-5.6-solmedium reasoning · full · 6 boards81.7%339 / 415 pts66.7%4 / 683.3%5 / 6211.3sper board181,094total03gpt-5.5high reasoning · full · 6 boards81.4%338 / 415 pts66.7%4 / 683.3%5 / 6605.7sper board295,640total04deepseek/deepseek-v4-pro-0813high reasoning · full · 6 boards77.6%322 / 415 pts66.7%4 / 683.3%5 / 6752.2sper board434,943total05gpt-5.6-sollow reasoning · full · 6 boards76.4%317 / 415 pts33.3%2 / 683.3%5 / 6101.0sper board151,049total06gpt-5.6-lunamedium reasoning · full · 6 boards63.4%263 / 415 pts50.0%3 / 666.7%4 / 6129.9sper board158,100total07gpt-5.5medium reasoning · full · 6 boards62.7%260 / 415 pts33.3%2 / 666.7%4 / 6173.7sper board174,017total08deepseek/deepseek-v4-flash-0731high reasoning · full · 6 boards61.7%256 / 415 pts33.3%2 / 666.7%4 / 6523.0sper board272,716total09deepseek/deepseek-v4-prohigh reasoning · full · 6 boards58.1%241 / 415 pts16.7%1 / 666.7%4 / 61655.6sper board368,005total10gpt-5.6-lunalow reasoning · full · 6 boards47.0%195 / 415 pts16.7%1 / 666.7%4 / 690.9sper board140,295total11gpt-5.5low reasoning · full · 6 boards20.0%83 / 415 pts0.0%0 / 666.7%4 / 683.3sper board138,490total12opencode/zai-coding-plan/glm-5.2low reasoning · full · 6 boards14.7%61 / 415 pts16.7%1 / 616.7%1 / 662.3sper board44,388total13opencode/zai-coding-plan/glm-5.1low reasoning · full · 6 boards0.0%0 / 415 pts0.0%0 / 60.0%0 / 60.0sper board0total14gpt-5.6-lunahigh reasoning · full · 6 boards0.0%0 / 415 pts0.0%0 / 60.0%0 / 6432.7sper board265,238totalA result counts as exact-optimal only when the model returns new tile placements as (x, y, z, letter) in one raw JSON tool envelope, with no candidates or solver hints.