DeepSeek V4.1 Flash

DeepSeek V4.1 Flash ran through OpenCode on DeepSeek’s own endpoint at high effort. DeepSeek offers low, high and max but no medium; the models in the October Anthropic, OpenAI, Google, xAI and Meta campaigns all ran at medium.

Read, then measure

On every seed it started with the manual and kept going back: 91, 83 and 91 documentation and research reads, against 24 to 35 for Opus and 25 for Sol on the same seeds. After each program ran it worked out a rate before deciding anything. On seed 1, after its first tree farm:

That habit made it good at the mechanics that need an algorithm. Its seed 1 cactus program splits the field between four drones, and each drone bubble-sorts its rows and columns four times before harvesting. On the first full run it found “full field sorted worked!”: 20,736 cactus a cycle.

def sort_rows_band(y0, band):
	n = get_world_size()
	for r in range(y0, y0 + band):
		for p in range(n):
			goto(0, r)
			for k in range(n - 1):
				if measure() > measure(East):
					swap(East)
				move(East)

def sort_cols_band(x0, band):
	n = get_world_size()
	for c in range(x0, x0 + band):
		for p in range(n):
			goto(c, 0)
			for k in range(n - 1):
				if measure() > measure(North):
					swap(North)
				move(North)

One cheap upgrade

The game hides some pages until you buy the upgrade they describe. The full fertilizer page, which explains that fertilized plants turn partly into weird substance, is one of them. Weird substance makes mazes, mazes pay gold, and gold buys drones. When the model bought fertilizer settled the rest of each seed.

On seed 1 it bought fertilizer at 27 minutes and read the page straight away. Weird substance followed at 54 minutes, mazes and gold at 55, and its first extra drone at an hour.

On seed 2 it couldn’t see what fertilizer was for:

It bought it at 1h41m, read the page, and found the chain at once: “Weird substance from fertilizing plants”. Gold came at 2h46m.

On seed 3 it read only the one-line cost page, searched for weird substance twice, and found nothing it could read:

The same thought ends “Skip mazes.” Seed 3 never made weird substance, gold or fertilizer.

Seed 1: seven workers

With gold flowing, seed 1 bought drones in steps, reaching eight at 2h16m. It also bought dinosaurs and hit a snag every seed would hit: every coloured hat was locked, and it found no way to take the dinosaur hat off.

Its answer was to leave the hatted drone out of the work. mega8b cashes in the tail, spawns seven workers that each farm one or two rows of carrots, trees and grass, and parks the main drone doing flips.

def worker(y0, cnt):
	n = get_world_size()
	while True:
		for i in range(cnt):
			goto(0, y0 + i)
			for j in range(n):
				do_tile()
				if j < n - 1:
					move(East)

change_hat(Hats.Dinosaur_Hat)
spawn_drone(worker, 0, 2)
spawn_drone(worker, 2, 2)
spawn_drone(worker, 4, 2)
spawn_drone(worker, 6, 2)
spawn_drone(worker, 8, 2)
spawn_drone(worker, 10, 1)
spawn_drone(worker, 11, 1)
goto(0, 0)
while True:
	do_a_flip()

Its first two minutes made 576 wood a second. From 3h27m it ran in half-hour stretches, broken by short sunflower and pumpkin runs, and each stretch added about three million wood.

That put seed 1 close to Grok 4.7, which also flew eight drones. Grok made more wood, carrots, pumpkins, weird substance and bones; DeepSeek more hay, cactus and gold.

Every model on seed 1

Seed 2: gold and one drone

Seed 2 reached mazes late, but its maze solver followed the right-hand wall and worked first time, at 14.4 gold a second. By 4h26m it had 51,840 gold. It read the Megafarm entry and liked it (“Very worth unlocking if we have time to use”, step 361), then changed its mind:

Seed 2 played all six hours on one drone. Its largest single haul came at the end, when it left a single-drone cactus sort running for the last 53 minutes and made 1.2 million cactus.

Seed 3: the bone loop

Without gold, seed 3 put its effort into dinosaurs. After trying every coloured hat it found that Hats.Straw_Hat takes the dinosaur hat off, which seeds 1 and 2 never found. Then it wrote boneloop: farm until there are 300 cactus to buy apples with, clear the field, and walk the dinosaur round the whole board.

while True:
    while num_items(Items.Cactus) < 300:
        for i in range(get_world_size()):
            for j in range(get_world_size()):
                crop(get_pos_x(), get_pos_y())
                move(East)
            move(North)
    for i in range(get_world_size()):
        for j in range(get_world_size()):
            if get_entity_type() != None:
                harvest()
                if get_entity_type() != None:
                    harvest()
            if get_ground_type() != Grounds.Soil:
                till()
            move(East)
        move(North)
    w = get_world_size()
    change_hat(Hats.Dinosaur_Hat)
    alive = True
    while alive and num_items(Items.Cactus) >= 2:
        x = get_pos_x()
        y = get_pos_y()
        if x == 0:
            if y == 0:
                alive = move(East)
            else:
                alive = move(South)
        elif y == 0:
            if x < w - 1:
                alive = move(East)
            else:
                alive = move(North)
        elif y % 2 == 0:
            if x < w - 1:
                alive = move(East)
            else:
                alive = move(North)
        else:
            if y == w - 1:
                if x > 0:
                    alive = move(West)
                else:
                    alive = move(South)
            else:
                if x > 1:
                    alive = move(West)
                else:
                    alive = move(North)
    change_hat(Hats.Straw_Hat)

Each run paid 20,449 bones, and the loop paid for its own apples. It ran for the last two and a half hours and made about 450,000 bones. At 4h31m the model listed what was missing:

What decided the score

The score adds up the logarithm of each resource’s production: a resource never made adds nothing, and each tenfold increase in one already made adds a single point. Seed 1 produced all eleven, with wood and hay above ten million. Seed 2 also opened all eleven, but on one drone its crops stayed between 70,000 and 230,000, and its 1.2 million cactus did most of the work. Seed 3 opened eight. Its 500,000 bones were worth 5.7 points, one more than 50,000 would have been, while gold, weird substance and fertilizer stayed at zero.

Same model, same settings: the seeds came down to when it bought one cheap upgrade and whether it trusted itself to use drones. It was also cheap to run, at $0.09 to $0.15 a seed at list price.