GPT-6.1 Sol beat Pokemon Red's first gym in 244 turns, a new record on PokeBench, for $2.60
GPT-6.1 Sol set a new record on PokeBench by defeating Brock in Pokémon Red's first gym in 244 turns. This achievement, costing $2.60, surpassed the previous record of 246 turns held by GPT-6 Astra, which cost $13.49. The run was conducted with a fresh save and a 1000-turn budget, with milestones read directly from game memory, ensuring an objective evaluation without human or model judgment.
时间与来源
时间显示为 UTC
显示时区:UTC
本地时区尚不可用,暂时显示 UTC。
发布当时偏移:UTC+02026年9月30日 00:56 UTC
收录当时偏移:UTC+02026年9月30日 06:00 UTC
- 发布
- 2026年9月30日 00:56
- 收录
- 2026年9月30日 06:00
- 来源类型
- 开发者社区
- 档位
- 社区
- 信源状态
- 正常
档位是按信源手工设定的编辑判断,不是逐条打分。
Same prompt, same harness, fresh save, 1000 turn budget. Milestones are read from game memory, so no human or model judges the run.
- Beat Brock on turn 244 (previous record: GPT-6 Astra, 246, $13.49)
- Wall time: 54 minutes, fastest on the board
- Cost: $2.60 at OpenAI's list price
- GPT-6 Sol, released one week earlier: 673 turns
Best moment: on turn 14 it picked up Bulbasaur by mistake, read the preview ("The rightmost ball contains Bulbasaur, not Squirtle"), and put it back.
Full turn logs and the leaderboard: https://pokebench.tv