Long-horizon play

Emerald Bench

Human controls · v2

Overall score

7 of 7 models · emerald-score-v2

Progress

7 of 7 models · star = first Hall of Fame reach · scored on active hours · axes end just past the slowest finisher

Breakdowns

7 of 7 models · highest or average level in the party over the retained playthrough

Leaderboard

7 of 7 models · best-ever milestones determine scoring

#	Model	Overall	Milestones	Active	Wall	Inference	Steps	Party	Levels	Badges	Pokédex	Credits
01	gpt-6-astra (high)	133.76	60/60	11h 52m	23h 19m	43%	6,895	5/6	45.6 / 73	8	9/127	90,767
02	claude-opus-5.5 (high)	104.62	60/60	30h 48m	55h 7m	41%	9,563	5/6	35.2 / 77	8	8/131	72,177
03	claude-fable-5.1 (high)	52.50	45/60	—	—	45%	—	3/6	34.0 / 57	7	5/97	141,755
04	gpt-6-sol (high)	43.17	37/60	—	—	58%	—	4/6	30.0 / 51	5	7/100	72,485
05	gpt-5.6-sol (high)	29.17	25/60	—	—	47%	—	3/6	21.7 / 37	3	4/66	106,752
06	claude-opus-5 (high)	24.50	21/60	—	—	42%	—	5/6	17.8 / 42	3	6/61	94,042
07	gpt-6-luna (xhigh)	22.17	19/60	—	—	66%	—	3/6	19.3 / 27	2	4/43	4,128

Score vs. cost

7 of 7 models · upper left is better