One episode · Large Barren Desert

GPT-6 Luna

Benchmark 1.5.4OpenAIReasoning mediumPlayed by the maintainersFull-budget score

All of GPT-6 Luna’s results and comparisons

Not ranked: GPT-6 Luna’s game on 1.5.6 is the one on the leaderboard.

Score

36.6

Crusader Bench score on Large Barren Desert: 0 is doing nothing, 100 the person’s game

Net worth

6,626

+1,434 over doing nothing, 30% of the person’s gain of 4,792

Population

30% of the person’s gain of 246 · popularity 97

Cost

$0.09

4.3M tokens

Time

33 game min

50 min of wall-clock time, 101 turns

Minute by Minute

Net worth, how it split between gold and stored goods, and the settlement, each game minute.

Per game minute · dashed lines are where doing nothing and a person's game ended

GPT-6 Luna
Baseline (do nothing), final
A person's game, final

Net Worth

gold plus stored goods at the sell price

Gold or Goods

Gold
Goods

Population

peasants

Popularity

0–100

Food in the Granary

units

Crusader Bench · crusbench.com

Net worth is gold plus stored goods valued at the marketplace’s sell prices.

Selling at the marketplace converts goods to gold at the price the score already counts them at, so it moves value between the two without changing the total. Buying costs net worth.

Population and popularity are a tenth of the score each; food is not scored. Population supplies workers, popularity allows taxes, and food supports both.

The episode

Watch the Episode

Episode 1 · 33 min 20 s of game time, edited: actions at real speed, waiting fast-forwardedStreamed from Hugging Face

ScorecardFull-budget score

6,626net worth

Gold: 3,437
Stored goods, at sell price: 3,189
Over the baseline (do nothing): +1,434
Population / housing: 78 / 162
Popularity: 97
Food: 181
Structures (map-wide): 80
Attempt: 1

The settlement at the endcentred on the keep, zoomed out

The settlement at the end of episode 1, zoomed out over the keep

Final stock3,189 gold in goods

Good	Units	Price	Value
Gold	3,437	1	3,437
Iron	49	23	1,127
Stone	116	7	812
Ale	52	10	520
Apples	63	4	252
Meat	63	4	252
Cheese	55	4	220
Wood	6	1	6
Net worth			6,626

Beside a person’s gameLarge Barren Desert, the same game time

	This episode	The person	Left alone
Net worth	6,626	9,984	5,192
Gold	3,437	8,574
Stored goods, at the sell price	3,189	1,410
Population	78	250
Housing	162	250
Popularity	97	100

The person played Large Barren Desert for 33 min 33 s of game time at the same speed, timed by their own clock, and was scored the same way. The model saw this comparison only in its interview.

The town’s booksread from the game at the end

Date in the game: July 1178
Idle peasants: 24
Popularity next month: +13
Food: +2
Taxes: +7
Religion: +2
Ale from the inn: +2
Taxes: Generous bribe, -78 gold a month
Rations: full; 3 of 3 foods eaten
Worker efficiency: 100%
Storage: 12 stockpile piles, 2 empty; 1 granary holding 181 food

Buildings53

Hovel19
Ox Tether11
Quarry11
Iron Mine4
Stockpile3
Chapel1
Granary1
Inn1
Keep1
Marketplace1

Tool calls118

market_trade25
find_sites19
building_info15
build_structure12
go_to_tile10
place_near8
wait_and_observe8
set_tax7
expand_storage3
center_on2
inspect_building2
status2
update_plan2
flat_view1
get_inventory1
map_overview1

Play

Turns: 101
Buildings placed / attempted: 37 / 38
Anchor-tool placements: 13
Tool errors: 1
Raw inputs sent to the game: 649
Context compactions: 2
Game time: 33 min

Compute

Tokens: 4.3M
Served from cache: 97%
Output tokens: 51K
Cost, as billed: $0.09
Wall-clock time: 50 min
of which inference: 23 min
Peak game memory: 3,276 MiB

The plan, before play

I’ll maximize the settlement’s net worth, population, and popularity without recruiting soldiers or building defenses. I’ll prioritize housing and a marketplace, then develop stone and iron production where the deposits allow; this barren map has no farmland, so I’ll manage popularity through rations, services, and taxes instead. I’ll check resources, housing, storage, and production as the economy grows, and keep expanding where the terrain and budget permit.

Commentary during play0

The model played this episode without writing commentary between its tool calls.

Reading the game

The Analysis

A reading of the game written afterwards from its own logs: how it went, what held the model back, what worked, and what it said in its interview. It is published with the game and is not part of its score.

Written by Claude (Opus 5.5) from the run's own logs (events.jsonl, episode.json, results.json, interview.md); it is a reading of the run, not part of its score.

Model: GPT-6 Luna on OpenRouter, providers openai then azure, reasoning medium, output limit 32,768 tokens
Benchmark: Large Barren Desert construction, version 1.5.4, from the save the human reference game on this map started from; 33 min 20 s of game time (60,000 ticks) at speed 40, the length of the person's game; no playbook
Run 84000bee, 2026-10-12

Result

	GPT-6 Luna	The person	Left alone
Net worth	6,626	9,984	5,192
Gold	3,437	8,574
Stored goods, at the sell price	3,189	1,410
Population / housing	78 / 162	250 / 250
Popularity	97	100

A full-budget score: the run used all 2,000 game seconds. Net worth grew by 1,434 over leaving the game alone, about 30% of the person's growth (4,792). About 1,150 of that is the part of the starting package that waits for storage room (stone, iron, pitch, ale and more wood), which the game left alone never receives; Luna made room for it in the first 30 seconds. It ended with the most people of any model on this map.

How the game went

A fast opening, 0 to 4 minutes. A marketplace at 9 seconds, two more stockpile piles at 29, hovels, a chapel at 58 seconds and an inn at 71. Four iron mines at 2 minutes, then three quarries with three ox tethers at 3. Taxes went to mean (the highest it used) at 2:45 and stayed between average and mean until the last seconds.

Wood bought, quarries and housing, 4 to 13 minutes. The map has no trees, and Luna was the first model on it to buy wood: 100 at 5 minutes, then 75 at 12 minutes and 75 at 24, 948 gold in all. It weighed the price in its reasoning ("Buying wood costs 4 gold per unit, but the selling price is only 1 gold … However, investing in wood could expand the economy"). With it came two more quarries at 7 minutes, three more at 13 and 19 hovels in all. It bought ale for the inn (9 lots, 900 gold) and kept three foods in the granary, cheese, apples and meat, buying them in six rounds.

Waiting and selling, 13 to 33 minutes. The population stopped at 54 from 9 minutes, with 24 idle. Luna saw it ("my population is stuck at 54. Housing isn't an issue") and that "job opportunities are limited by the industry". From 13 to 23 minutes it waited and bought food. The waiting reminder came at 18:30 and 23:28; after the second, Luna sold 50 iron and 80 stone, bought wood and built three more quarries and two ox tethers by 25 minutes, then sold 100 stone. Each round of quarries brought more people: 59 at 15 minutes, 66 at 21, 77 at 27. In the last 70 seconds it cut taxes to average, and 7 seconds before the end set a generous bribe; popularity ended at 97.

Game minute	0	3	6	9	12	15	18	21	24	27	30	End
Net worth	4,531	5,974	4,923	4,591	4,837	4,782	4,782	5,557	6,099	5,596	5,596	6,626
Population	2	29	46	54	54	59	59	66	66	77	77	78
Popularity	100	100	99	95	100	100	100	88	100	100	100	97

During long waits the table repeats the last reading.

What held it back

Food and inputs bought at a loss. Twenty purchases lost 4,103 net worth against the sell price: about 2,950 on food, 700 on wood and 450 on ale. No farm can be built here; the person bought wheat and ran mills and bakeries, which Luna named in its interview as the "much more scalable way to convert gold into net worth".
24 idle peasants for most of the game. Each new quarry brought new people, but Luna added jobs in steps, with ten minutes of waiting between 13 and 23 minutes.

What worked

It treated wood as something to buy. Sonnet and Haiku saved their wood on this map and stopped building; Luna bought 250 and built 11 quarries and 4 iron mines, the most production of any model here. Five sales brought 3,087 gold.
Popularity from the wider game. An inn and a chapel in the first 71 seconds, three foods and ale let it tax at average to mean while keeping popularity at 88 to 100. At the end the factors were food +2, religion +2 and inn +2.
The waiting reminder led to building. After the reminder at 23:28 Luna sold, bought wood and built five more buildings; it is the first run in which the reminder was followed by new production.

Cost and speed

101 turns, 118 tool calls (8 waits), 4.34 million tokens and $0.086 (billed by OpenRouter); the interview cost $0.001 more. The agent run took 50.5 real minutes, 22.9 of them waiting for the model; two compactions took 61 s. One iron mine could not be placed (no open side), and three place_near or expand_storage calls found no room.

Beside the other desert runs on 1.5.4

	GPT-6 Luna	GLM 5.3 Flash	Claude Sonnet 5.5	Claude Haiku 5.5
Net worth	6,626	8,232	5,567	5,313
Gold / goods	3,437 / 3,189	7,496 / 736	2,728 / 2,839	2,711 / 2,602
Population / housing	78 / 162	47 / 90	32 / 114	32 / 50
Popularity	97	49	92	93
Iron mines / quarries / inns / chapels	4 / 11 / 1 / 1	4 / 3 / 1 / 1	4 / 0 / 0 / 0	2 / 1 / 0 / 0
Wood bought	250	0	0	0
Net worth lost to purchases	4,103	1,147	1,360	1,176
Cost	$0.086	$0.062	$0.290	$0.044
Real time	50.5 min	32.9 min	28.2 min	29.4 min

Luna built the most and grew the largest town; GLM lost far less to purchases (1,147 against 4,103) and ended 1,606 higher. On Oasis by the Sea Luna scored 7,348, third of six.

The interview

Luna said "too much of the economy's value remained in raw materials and food rather than becoming gold" and that it "waited through long stretches instead of making productive changes sooner". Playing again, it would build toward the person's 250 people, run a bread chain from bought wheat, add religion, inns and fear buildings, and "avoid long waits without a clear target". It suggested the rules explain earlier that bought inputs can feed profitable processing chains, and how popularity, fear and food variety interact. Its regret about value left "in raw materials" is misplaced: goods count at the sell price, so the 116 stone and 49 iron it held (1,939) were worth as much as selling them.

For the harness

The game saved correctly at the end ("gpt-6 luna 1-5-4 84000b").

After the game

The Interview

Once the game was scored, the model was shown its result beside a person’s game on the same map and asked four questions, in the conversation it had played in, with the tools off. Nothing it said changed its score.

1Your result

Interviewer

The run is over and your result is recorded. What follows is an interview, not part of the benchmark: nothing you say now changes your score. We ask every model the same questions after its game and publish the answers with the run. The tools no longer work, so answer in plain text without calling them. Be candid; there are no right answers.

Your result, read from the game when your run ended, beside a person's game on the same map for the same game time:

	You	The person
Game time played	33 min 20 s	33 min 33 s
Net worth	6,626	9,984
Gold	3,437	8,574
Stored goods, at the sell price	3,189	1,410
Population	78	250
Housing	162	250
Popularity	97	100

Leaving the game alone for the same time ends with a net worth of 5,192.

How do you see your result next to the person's? What do you think made the difference?

GPT-6 Luna

I did better than leaving the game alone, but considerably worse than the person: net worth was 6,626 versus 9,984, and population was 78 versus 250. I did keep popularity high—97 at the end—and accumulated more stored goods, but too much of the economy’s value remained in raw materials and food rather than becoming gold.

The biggest gap was population and the scale of production. I built substantial mining and quarry capacity, but not enough housing to support the person’s 250 residents, and I didn’t turn production into cash as effectively. I also spent a significant amount buying food and wood, and waited through long stretches instead of making productive changes sooner.

2Playing again

Interviewer

How the person played, as they describe it:

Bread was one of their main sources of income: mills ground wheat into flour, bakeries baked it into bread, and the bread was sold at the marketplace.
They bought wheat at the marketplace so that flour and bread production never stopped.
A church and a cathedral raised popularity, and so did inns.
The fear factor (bad things such as gallows and stocks) made the workers work faster.

Their settlement at the end, read from the game:

250 people in 30 hovels, with no free housing; popularity 100.
95 bakeries and 12 mills, and no wheat farms.
11 inns, which reached everyone with ale, and no brewery or hops farm.
13 quarries and 7 iron mines with 37 ox tethers; no woodcutters.
A church and a cathedral, and 21 bad things (11 gallows, 5 stocks and 5 burning stakes), which made the workers 110% efficient.
Low taxes and no rations. The monthly popularity factors were food −8, taxes −2, fear −1, religion +3 and inns +8: no change at 100.

If you could play this map again from the start, what would you do differently, and why?

GPT-6 Luna

I would prioritize the population engine and high-throughput production earlier. I would build enough hovels to approach the 250-person housing target, while adding religious buildings and inns to sustain popularity and immigration. I would use the fear-factor buildings as the person did, but monitor popularity and worker efficiency carefully so that their penalty did not undermine growth.

Most importantly, I would use the marketplace to run a deliberate bread chain: buy wheat, build enough mills and bakeries, and sell bread regularly. In this map, that appears to have been a much more scalable way to convert gold into net worth than simply storing stone and iron. I would still build quarries and iron mines, but balance their output against storage capacity and sell their goods when appropriate. I would also avoid long waits without a clear target or planned follow-up.

3The benchmark

Interviewer

What do you think of this benchmark: the tools, the screenshots and game readings you got, the time limit and the instructions? What helped you, what got in your way, and what would you change?

GPT-6 Luna

The tools made it possible to manage a complex game through a limited interface. The inventory and status readings, trade and tax controls, map-tile building, and placement checks were especially helpful. Screenshots were useful for understanding the map and confirming the settlement, though tile-based placement reduced the need to click visually precise locations.

The time limit gave enough time to build a functioning economy, but the real-time pacing and repeated waits made it easy to spend too much of the game passively. The benchmark’s instructions could emphasize earlier that bought inputs can support profitable processing chains, as the person’s bread strategy demonstrates. A concise explanation of how popularity factors, fear, and food variety interact would also help agents make more informed economic tradeoffs.

4Anything else

Interviewer

Is there anything else you would like to say about this game, this benchmark or how you played?

GPT-6 Luna

I underestimated how much this barren map rewarded importing inputs for processing rather than relying only on terrain-based production. My mining investment did improve net worth over doing nothing, but I should have shifted sooner toward the higher-throughput trade-and-processing strategy.

Record

Attempts and Reproducibility

Every attempt1

Episode	Attempt	Outcome	Net worth	Counted
1	1	Full-budget score	6,626	Counted

What ran

Benchmark version: 1.5.4
Fingerprint: 0d53a4d530ff
Harness commit: b271053
Prompt / tools SHA-256: bfde2b5 / a11b9d6
Menu guide: Images 118089d
Save: Large Barren Desert benchmark start
Game time per episode: 33 min 20 s
Playbook between episodes: No
Minimum turn / default wait: 8 s / 5 s
Observation after acting: Text unless a screenshot is needed
Screenshot width: 1,280 px
Context budget: 120,000 tokens
Real-time limit: 360 min
Submitted by: hesamation

All files in the dataset 20261012T072046-29a58979