GPT-6 Astra

A familiar first five minutes

Astra’s openings were almost a copy of Sol’s. On all three seeds its first wood program harvested every bush the moment it planted it, so nothing ever grew. The fix was refused because the program was still running, and the old version ran for three more minutes without a single log before Astra wrote a second program, wood2, that waits until a bush is ready. Sol’s write-up tells the same story; Astra repeated it on seed 2 and seed 3, losing about six minutes each time.

That second program set a habit that lasted all game. Astra wrote 17 to 20 programs per seed and never edited one: every change got a new name. Seed 2 ran economy, economy2 and economy3, then supplies and supplies16; its gold programs were gold, gold3 and finally goldfast.

Everything on a timer

From about an hour in, all three runs follow the same rhythm. Astra runs one program for a fixed stretch, usually ten, fifteen or twenty game-minutes, checks what it earned, buys upgrades, and moves on to the next: a crop farm for hay, wood and carrots, then pumpkins, cactus, gold maze runs and dinosaur bones, then back to the crop farm. On seed 3 the cycle starts at 1h09m with a pumpkin program, and by 1h40m it has been through pumpkins, sorted cactus and gold.

It’s an orderly way to open everything, and it worked: bones, the last resource on every seed, arrived by 3h45m at the latest. The cost shows later. A specialist’s first stint is worth four to six points, because it opens a new resource; repeating it mostly adds one or less. Seed 2’s three later pumpkin stints added 1.4, 0.9 and 0.5 points, while its crop-farm stints in the same hours added 2.6 to 3.6. Yet in the last four hours the crop farm got about 100 minutes on seeds 1 and 3 and under 80 on seed 2, and the rest went round the specialists. Score is not everything a stint bought, though: pumpkins paid for field expansions and cactus for dinosaur and maze upgrades.

The hat it forgot to take off

To farm bones, a drone puts on a dinosaur hat and drags a growing tail around the field; the bones are paid out when the hat comes off. Astra’s seed 1 bones program walks the field in a loop that never crosses itself, and only swaps back to the straw hat once the dinosaur can’t move:

while True:
	change_hat(Hats.Dinosaur_Hat)
	moving = True
	while moving:
		x = get_pos_x()
		y = get_pos_y()
		if y == 0:
			if x == 0:
				moving = move(North)
			else:
				moving = move(West)
		elif x % 2 == 0:
			if y == n - 1:
				moving = move(East)
			else:
				moving = move(North)
		else:
			if y == 1:
				if x == n - 1:
					moving = move(South)
				else:
					moving = move(East)
			else:
				moving = move(South)
	change_hat(Hats.Straw_Hat)

Stopped by the clock in the middle of that loop, it left the hat on. The next program was the pumpkin farmer, which never touches hats, and it spent twenty minutes farming pumpkins with the dinosaur hat still on: 2,960 pumpkins, against 79,872 in the fifteen-minute pumpkin stint an hour earlier. The tail was cashed in only during the next program, a cactus sorter that starts by clearing the field: 11,250 bones turned up in the middle of a cactus run.

Before its next program, Astra’s reported thinking was just “Clearing the hat”, and from then on its programs started with exactly that: pumpkin_fast opens with change_hat(Hats.Straw_Hat), and on seed 3 it wrote a two-line program just to take the hat off after each bones run.

The payout rule also explains two bones stints late on seeds 2 and 3 that look empty in the step log. On seed 3, fifteen minutes of bones show no bones at all; the next action, reset_hat, banks 258,064. On seed 2 a twenty-minute run shows nothing and the following gold program, which starts by putting the straw hat back on, collects 1,486,088. The dinosaurs were working the whole time.

Drones bought, not used

The game lets you buy extra drones, and Astra did on every seed. What it did with them differed.

On seed 3 the crop farm is industry, written once at 1h55m. It farms columns of grass, trees with bushes between them, and carrots with a row of sunflowers, fertilising every tree, which is where seed 3’s weird substance came from. Its last lines start the workers:

for i in range(1,4):
	spawn_drone(work, i * get_world_size() // 4)
work(0)

Four drones, fixed. At 3h37m Astra bought its way to eight drones and ran industry straight after; the game’s telemetry shows four drones flying through that stint and the two that followed. Only its cactus sorter, which starts a drone per row, ever put all eight in the air. Seed 2 went the same way: eight drones bought at 4h13m, then the same four-drone supplies16, with all eight working only in the last seventeen minutes.

Seed 1 is the exception that shows what the drones were worth. Astra bought eight and then sixteen drones together at 5h20m and immediately wrote a crop program for sixteen, which made 66.8 million wood in the last 39 minutes, more than in the whole game before it.

At the end of seed 3 Astra was also holding 147,196 gold, more than the 128,000 that took seed 1 to sixteen drones. Seeds 2 and 3 spent their late pumpkins on a bigger field instead, 22×22 instead of 16×16, and mostly worked it with the same four drones.

What decided the score

Each resource scores about one point per tenfold of production, so the crops decide most of it. Astra’s best seed, seed 3 (69.91), had 100 million hay, 253 million wood and that 115 million weird substance. Its worst, seed 2 (63.16), never got its crops much past a few million: 3.4 million hay and 8.8 million wood, where Sol made 2.2 billion wood on the same seed. Hay, wood and carrots alone account for nearly seven of the 12.7 points between them.

Every model on seed 2

Where Astra did well was breadth. Every resource was open on every seed, and bones in particular went better than for Sol: 1.9 million to 101,000 on seed 1. But one point per tenfold rewards depth in the crops more than breadth, and that left Astra third on every seed, behind Opus and Sol, at $1.73 to $1.88 a run.