Kimi K3.1 Pro

Kimi K3.1 Pro ran in Kimi Code, Moonshot’s own agent, at high effort. Moonshot offers low, high and max but no medium; the models in the October Anthropic, OpenAI, Google, xAI and Meta campaigns all ran at medium.

Batches, not a layout

The first half hour looked like most models’: a hay loop, then bushes that didn’t grow at first because the program got to them too early, then a mixed field of grass, trees and carrots. At 41 minutes on seed 1 it settled on the shape it would keep on every seed:

Rather than give each crop a corner, Kimi alternated whole-field programs: one to grow carrots, hay and wood, then one to turn the whole field into pumpkins. On seed 3 that’s mixed and pumpkin_multi; on seed 2, mega_mixed and mega_pumpkin. It took seed 1 half an hour to make the switch work. The first ten minutes brought 115 pumpkins, the next ten two, and only after three fixes did it report about 2,000 per ten minutes.

The hat that comes off

Bones come from a dinosaur that grows a tail by eating apples, paid out when the dinosaur hat comes off. Taking it off means putting another hat on, and all three seeds reached for the same two first, green and gray, and found both locked (seed 1, seed 2, seed 3).

Seed 3 tried a third:

It worked, and the next run paid two full tails of 1,225 bones each. Seed 1 had the same idea, twice, and didn’t act on it:

It tried brown and purple instead, both locked. Meanwhile the dinosaur hat stayed on after its program stopped, and a drone in that hat can’t wrap round the field’s edges. The crop program, which relies on wrapping, ran from 2h57m to 3h29m and added half a point while Kimi puzzled over missing pumpkins, until it found the cause:

Its way out was to spawn a fresh drone to do the farming and let the hatted one finish. Seed 2 tried the dinosaur once, judged it inefficient without a way to unequip, and never came back. It ended with 29 bones.

Seed 1: four drones, one worker

Seed 1 bought the first drone upgrade at 2h33m and the second, for four drones, at 4h21m:

No program after that split the work. main was last revised at 3h29m, and its one spawn_drone call is the hat workaround: a single worker running the whole field. The game’s telemetry shows one drone flying for the whole run. Its closing summary still listed “Four-drone capacity”. With one drone, seed 1’s crops stayed in the thousands to tens of thousands: 11,721 hay, 28,300 wood, 5,695 carrots.

Seed 2: wood by the million

Seed 2 went the other way. From 2h44m it ran four drones, and it put almost everything into growing the field: 12×12 at 2h56m, 16×16 at 3h25m, and, after a long pumpkin drive for the 512,000 the upgrade cost, 22×22 at 5h29m. In its last half hour the four workers of mega_mixed22 made 1.4 million wood in ten minutes.

The detour that didn’t pay came in the middle. At 3h32m it set out to buy eight drones with gold from 16×16 mazes, but its depth-first solver managed seven mazes in fifteen minutes. It dropped the plan for a sorted 16×16 cactus field:

cactus16 bubble-sorts the field, pass after pass, until a pass makes no swaps. After twenty minutes it hadn’t finished; a 12×12 retry ran another twelve before Kimi stopped it. Those 54 minutes added half a point, and cactus stayed at the 4,096 of its first harvest.

Seed 3: one harvest, five points

Seed 3 reached 16×16 at 4h42m and tried the same pair of big harvests. Its first sort did a fixed number of passes and harvested a field that wasn’t sorted yet, for 412 cactus. It went back to sorting until nothing moves, the same loop as seed 2’s:

def sort_field(size):
	changed = True
	while changed:
		changed = False
		for y in range(size):
			for x in range(size):
				if x < size - 1 and measure() > measure(East):
					swap(East)
					changed = True
				if y < size - 1 and measure() > measure(North):
					swap(North)
					changed = True
				move(East)
			move(North)

This time it finished: 65,536 cactus within ten minutes. The record doesn’t show why the same idea stalled on seed 2. With apples to spare, seed 3 then scaled its dinosaur route to the full field, and costed it first:

A tail that fills 255 of the 256 squares, squared, and doubled by the dinosaur upgrade it had just bought, is 130,050 bones, and that is what the run paid.

change_hat(Hats.Straw_Hat)
go_home()
size = get_world_size()
for y in range(size):
	for x in range(size):
		if can_harvest():
			harvest()
		move(East)
	move(North)
go_home()
change_hat(Hats.Dinosaur_Hat)
running = True
while running:
	if not travel(East, size - 1) or not travel(North, size - 1):
		running = False
		break
	up = False
	for x in range(size - 2, 0, -1):
		if not travel(West, 1):
			running = False
			break
		if up:
			if not travel(North, size - 2):
				running = False
				break
		else:
			if not travel(South, size - 2):
				running = False
				break
		up = not up
	if not running:
		break
	if not travel(West, 1) or not travel(South, size - 1):
		running = False
		break
change_hat(Hats.Straw_Hat)

That one harvest was worth 5.1 points. With gold from 16×16 mazes it bought eight drones for the last half hour.

What decided the score

The score adds up the logarithm of each resource’s production, so each tenfold increase in one resource adds a point, and a resource barely made drags the total down. Seed 2 scored highest, 52.79, from crops in the hundreds of thousands to millions, with 29 bones and 4,096 cactus holding it back. Seed 3, at 49.17, had nothing above 205,000, but its six biggest resources, from cactus and bones to wood, all landed between 114,000 and 205,000; power, at 135, was its weak spot. Seed 1, at 42.38, opened everything on one drone and stopped there.

The same seeds show how differently a similar total can be built. On seed 2, DeepSeek V4.1 Flash scored 51.07 from a farm shaped the opposite way: 229,000 wood against Kimi’s 10.4 million, but 1.29 million cactus and 33,668 bones.

Every model on seed 2

On seed 3 Kimi and Sonnet finished level, 49.17 each. Both scores rest on one big dinosaur harvest, Sonnet’s more than ten times larger; Kimi’s crops were larger, with 204,000 wood to Sonnet’s 120,000 and 114,000 cactus to 37,000.

Every model on seed 3

The runs took 13 to 19 minutes of real time and $1.63 to $3.02 a seed at Moonshot’s list prices; they ran on a Kimi Code subscription.