Grok 4.7

Knew the game, read the manual anyway

Grok opened every seed by saying it already knew The Farmer Was Replaced: “I know this game well” on seed 1, and the same words on seed 2 and seed 3. Then it studied harder than any other model: 102 documentation and research reads on seed 1 and 83 on seed 2, the most on those seeds, and 65 on seed 3, level with Fable. Opus made 24 to 35.

Some of that prior knowledge was off. On seed 3 it found the pumpkin yield “differs from what I remembered”. And on seed 1, in its first minute of planning, it guessed something about the field that would matter all game:

The field goes to 32×32, as Opus and Sol both showed.

A small farm on purpose

The early game went quickly. Pumpkins were planted at 34 minutes on seed 1, gold came out of mazes before the hour on seeds 1 and 2, and all eleven resources were flowing by 2h53m at the latest. What Grok didn’t do was grow the field.

On seed 3, at 1h03m, it worked out that its single drone took about 4.7 seconds to sweep a 6×6 pumpkin patch that regrew in under a second, and drew a conclusion:

It did get drones, four by 1h58m, but only expanded to 8×8 at 3h17m, for 1,000 pumpkins. On seed 2 it costed out the next size up at 1h28m and liked what it saw:

The same thought ends with “Expanding to 12x12 is too risky with only two drones”. The upgrade needed 8,000 pumpkins, and by then Grok had stopped growing them. Seed 2 stayed at 8×8 to the end. Seed 1 reached 12×12 at 4h15m, with less than two hours left.

The reasoning is sound for a single drone. But each extra drone sweeps its own columns, so the leaders grew field and fleet together. Opus had 32 drones and a 32×32 field by 4h40m on every seed.

Gold, slowly

Drones cost gold, gold comes from mazes, and mazes turn out to be slow to solve. Grok said so itself at the end of seed 1:

Each Megafarm level doubles the drone count and quadruples the price. Grok reached eight drones on seed 1 at 1h53m, then left gold at 50,368 for the rest of the game; the next level cost 128,000. Seed 3 got to eight at 3h15m. Seed 2 stopped at four.

Every model on seed 1

A dinosaur on every seed

Bones come from a dinosaur that grows a tail by eating apples, paid out when the hat comes off. On every seed Grok’s first attempts made almost nothing. On seed 1 at 2h57m:

It took nine rewrites in eight minutes on seed 1, starting from a fixed path round the field, with tests on shrunken 2×2 and 3×3 fields along the way. Seeds 2 and 3 went through the same discovery: a three-segment tail is worth nine bones. The version that worked was simple: chase each apple directly, cash in after forty, and repeat. On seed 1 it made 55 bones a second, enough to buy Polyculture’s next level twice.

def one_dinosaur():
    change_hat(Hats.Dinosaur_Hat)
    eaten = 0
    while eaten < 40:
        pos = measure()
        if pos == None:
            break
        tx, ty = pos
        guard = 0
        while (get_pos_x() != tx or get_pos_y() != ty) and guard < 25:
            guard = guard + 1
            if not step_toward(tx, ty):
                guard = 25
        if get_pos_x() != tx or get_pos_y() != ty:
            break
        eaten = eaten + 1
    change_hat(Hats.Straw_Hat)

Plants and their companions

Polyculture multiplies a crop’s yield when the neighbour it asks for is planted where it asks. Grok wrote for it on every seed, and it carried the endings. On seed 1 the crop field is a checkerboard of trees and carrots, and after each planting the drone walks to the square the plant asked about, plants its companion there, and walks back:

def do_tile():
    entity = get_entity_type()
    if can_harvest():
        if entity == Entities.Tree and num_items(Items.Fertilizer) > 0 and num_items(Items.Weird_Substance) < 800000:
            use_item(Items.Fertilizer)
        harvest()
        entity = None
    if entity == None or entity == Entities.Grass:
        if (get_pos_x() + get_pos_y()) % 2 == 0:
            plant(Entities.Tree)
        else:
            if get_ground_type() != Grounds.Soil:
                till()
            plant(Entities.Carrot)
    comp = get_companion()
    if comp == None:
        return
    wanted, pos = comp
    x, y = pos
    sx = get_pos_x()
    sy = get_pos_y()
    goto(x, y)
    if get_entity_type() != wanted:
        plant_wanted(wanted)
    goto(sx, sy)

def poly_column():
    for j in range(get_world_size()):
        do_tile()
        move(North)

def parallel_poly():
    for i in range(get_world_size()):
        if not spawn_drone(poly_column):
            poly_column()
        move(East)

That differs from Opus’s version, which writes the request down and plants the companion when it next passes that square. Grok’s costs extra walking, but it splits cleanly across drones: on seed 1 at 3h26m, one drone made about 612 wood a second and eight made about 1,595.

From 3h28m to 5h48m that program added 9.3 points on seed 1, nearly a sixth of the final score, and 33 million wood. The last long stint added 7.2 points on seed 2 and 10.3 on seed 3.

Seed 2 waits it out

Seed 2 is the lowest of the three, and its last three hours show why. After buying a Polyculture level at 2h51m it started its crop program and then mostly waited: ten tool calls in 189 minutes, including an hour and 58 minutes at a stretch, broken only to buy yield upgrades.

That program grew only crops, and four drones’ worth of them. Pumpkins had stood at 11,102 since 1h05m, gold at 11,008 since 1h38m, and bones stopped at 61,524. Seed 3, given the same final stretch, kept rotating through cactus, bones and mazes and ended with 574,209 cactus and 91,500 gold against seed 2’s 16,384 and 11,008.

What decided the score

Across all three seeds, wood, hay and carrots carried the score: 15 to 33 million wood, 3 to 4 million hay, up to 4 million carrots, about 20 points between them. Grok beat Sonnet on seeds 1 and 3: on seed 1 through bones, power and weird substance, on seed 3 by nearly nine points, mostly crops, cactus and power.

Every model on seed 3

Against the leaders the picture inverts. Opus and Sol produced hundreds of millions to billions of each crop on fields up to sixteen times larger, and Opus’s cactus, pumpkins and gold ran about two to four orders of magnitude beyond Grok’s. Every resource Grok made is in its score, and the small field capped all of them.

The runs were slow and expensive too: 49 to 70 minutes of real time and $9.31 to $13.92 a seed. Those cost figures come from Cursor’s own per-turn token totals, which can’t be broken down request by request, so they are less certain than the other models’.