One episode · Large Barren Desert

Gemini 4 Argon

Benchmark 1.5.6GoogleReasoning mediumPlayed by the maintainersFull-budget score

All of Gemini 4 Argon’s results and comparisons

Score

64.8

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

Net worth

8,253

+3,061 over doing nothing, 64% of the person’s gain of 4,792

Population

185

74% of the person’s gain of 246 · popularity 63

Cost

$0.84

1.7M tokens

Time

33 game min

39 min of wall-clock time, 58 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

Gemini 4 Argon
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

8,253net worth

Gold: 1,327
Stored goods, at sell price: 6,926
Over the baseline (do nothing): +3,061
Population / housing: 185 / 194
Popularity: 63
Food: 241
Structures (map-wide): 189
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 stock6,926 gold in goods

Good	Units	Price	Value
Gold	1,327	1	1,327
Stone	474	7	3,318
Iron	98	23	2,254
Bread	241	4	964
Pitch	20	10	200
Flour	10	10	100
Ale	7	10	70
Hops	2	8	16
Wood	4	1	4
Net worth			8,253

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

	This episode	The person	Left alone
Net worth	8,253	9,984	5,192
Gold	1,327	8,574
Stored goods, at the sell price	6,926	1,410
Population	185	250
Housing	194	250
Popularity	63	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: −8
Food: −4
Taxes: −12
Ale from the inn: +8
Taxes: Extortionate taxes, 258 gold a month
Rations: half; 1 of 1 foods eaten
Worker efficiency: 100%
Storage: 28 stockpile piles, 9 empty; 3 granaries holding 241 food

Buildings149

Ox Tether34
Bakery33
Hovel23
Brewery13
Quarry12
Inn7
Stockpile7
Water Pot6
Mill5
Iron Mine4
Granary3
Keep1
Marketplace1

Tool calls230

find_sites57
build_structure48
market_trade35
observe20
building_info14
center_on11
wait_and_observe10
go_to_tile8
set_tax7
expand_storage3
go_to_view3
place_near3
save_view3
set_rations3
update_plan3
demolish_building1
map_overview1

Play

Turns: 58
Buildings placed / attempted: 134 / 179
Anchor-tool placements: 13
Tool errors: 7
Raw inputs sent to the game: 1,070
Context compactions: 3
Game time: 33 min

Compute

Tokens: 1.7M
Served from cache: 84%
Output tokens: 16K
Cost, as billed: $0.84
Wall-clock time: 39 min
of which inference: 9 min
Peak game memory: 3,490 MiB

The plan, before play

Understanding of the Task

Scenario: Large Barren Desert Free Build (33.33 minutes of game time).
Objective: Maximize final net worth (gold + stored goods valued at marketplace sell prices: iron 23, pitch 10, ale 10, flour 10, wheat/hops 8, stone 7, food 4, wood 1), followed by population and popularity.
Constraints: No military recruitment or defensive structures. No oasis grass exists on the map, meaning farms cannot be built. Weapons/armory do not count toward net worth.
Map & Starting Resources: Granary and Stockpile already exist. The starting package (~5,000 gold, 55 meat, ~245 wood, 50 stone, 15 iron, 20 pitch, 20 ale) arrives initially over the first ~27 game seconds, filling the initial 4-pile stockpile with wood until more stockpiles are added. Iron ore strips lie ~60 tiles east of the keep; stone deposits lie ~100 tiles west.

Opening Plan

Survey & Expand Storage: Call map_overview (free read before acting) to locate exact coordinates for iron ore, stone deposits, and any trees/oil, then immediately expand the stockpile (expand_storage) so the rest of the starting package can be delivered, and build a Marketplace and initial Hovels.
Establish High-Value Raw Material Production: Focus heavily on Iron Mines (east, 23 value per unit, no oxen required) and Quarries with Ox Tethers (west, 7 value per unit), plus Woodcutters if trees are nearby or trade wood as needed.
Manage Population, Rations, and Taxes: Use rations, imported/processed food (e.g. flour/bread or ale/inn if profitable, or adjusting rations/taxes carefully), and housing to grow the workforce rapidly to staff all mines and quarries while keeping popularity above 50.
Scale Continuously: Keep expanding stockpiles, housing, and mines/quarries throughout the 33-minute budget so no worker or storage bottleneck halts net worth growth.

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: Gemini 4 Argon (google/gemini-4-argon) on OpenRouter, served only by Google AI Studio (no fallbacks), reasoning medium, output limit 32,768 tokens
Benchmark: Large Barren Desert construction, version 1.5.6, the version of this map's round, from the save the human reference game 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 956efdbe, series 20261013T191602-69ff6166, 2026-10-13. An earlier start on this map (series 20261013T175345-d78ce5c4) was stopped after two game minutes by the operator's cost guard, which took one uncached request for a caching failure; this series is a fresh start with no memory of it.

Result

	Gemini 4 Argon	The person	Left alone
Net worth	8,253	9,984	5,192
Gold	1,327	8,574
Stored goods, at the sell price	6,926	1,410
Population / housing	185 / 194	250 / 250
Popularity	63	100

A full-budget score: the run used all 2,000 game seconds. Net worth grew by 3,061 over leaving the game alone, 64% of the person's 4,792 and the second most on this map after Claude Opus 5.5's 3,565. Its 185 people are the most on this map (GPT-6 Luna's 106 come next); it ended with 24 idle. The stored goods are 474 stone (3,318), 98 iron (2,254), 241 bread (964), 20 pitch (200), 10 flour (100), 7 ale (70), 2 hops (16) and 4 wood (4).

How the game went

The opening, 0 to 6 minutes. Three more stockpile piles at the start, the marketplace at 0:17 and three hovels at 0:26, four iron mines at 0:48 and four quarries at 1:14, three more quarries at 2:18 and ox tethers in fours from 3:05, and eight hovels at 3:53 and 4:08. It set rations to none and taxes to low at 1:49, and back to full rations and no taxes at 3:21. 100 wood bought at 2:38, 4:27 and 5:47. Then the bread chain, on bought grain since this map has no farmland: 25 wheat and 25 flour at 4:57, two mills and a bakery at 5:19, and a bakery, two inns and a brewery at 5:31.

Bakeries and taxes, 6 to 24 minutes. Bakeries in fours from 7:26 (two batches at 7:43 and 8:07 were refused because their screenshot was too old), 50 wheat at 9:00 and then wheat in fives to twenties every few minutes, a twelfth quarry by 10:20, tethers in fours, hovels at 12:45 and 13:49, inns at 13:00, 14:04, 16:39, 19:54 and 23:12 (seven) with breweries beside them and hops bought, and two more granaries at 16:24. Taxes went to average at 9:16 and mean at 11:33, and stayed at mean or average with full rations; popularity held between 65 and 86 from 6 minutes, mostly at 70. Gold fell from 4,722 at 3 minutes to 716 at 12, and stayed under 1,100 from then until the last tax payment, as it spent each payment on wood and wheat.

The end, 24 to 33 minutes. Bakeries to 33, breweries to 13, mills to five, tethers to 34 and hovels to 23; from 28:41 six water pots, which give idle peasants jobs. In the last two minutes it waited for gold instead of buying. At 32:48 it set half rations and at 32:53 extortionate taxes; popularity fell from 70 to 63 in the remaining 25 seconds, and the last payment, at 33:15, brought 249 gold.

Game minute	0	3	6	9	12	15	18	21	24	27	30	End
Net worth	4,711	5,688	3,516	3,321	2,331	2,209	2,529	3,313	4,125	4,694	5,535	8,253
Gold	4,440	4,722	2,531	2,271	716	361	279	429	279	158	74	1,327
Iron	0	15	22	28	37	45	53	63	71	81	89	98
Stone	0	50	50	50	74	74	98	138	218	258	362	474
Food	48	48	22	0	45	69	69	115	160	207	214	241
Population	2	23	35	54	64	83	95	118	141	151	170	185
Popularity	100	75	70	77	84	69	70	70	70	70	70	63

Up to 30 minutes the net worth row counts gold, wood, stone, iron and food (at 4) as the recording logs them; pitch, flour, ale and hops are left out, so it reads a little low. Gold in the 30-minute column is a low point between payments.

What held it back

Value bought, not made. 10,090 gold went on wood, wheat, hops and flour, 6,900 more than they are worth at the sell price, and nothing was sold. Net worth, as the recording counts it, was below the starting 4,711 from about 5 minutes to about 26.
Bread eaten. Full rations for up to 185 people from 3:21 to 32:48; the granaries held 241 loaves at the end. The person ran no rations and low taxes, held at 100 popularity by inns, a church and a cathedral, and sold the bread; Gemini's interview says the same.
Population stopped at 24 idle, the limit the 1.5.6 rules describe, with 9 places still free; the water pots came at 28:41.

What worked

The bread chain on bought grain: 33 bakeries and five mills by the end, which fed the largest population on this map.
Mean taxes for most of the game with popularity near 70, held by seven inns.
Twelve quarries and the map's four iron mines from the first three minutes: 474 stone and 98 iron at the end, 5,572 of the net worth.

Market

1,100 wood, 180 wheat, 50 hops and 25 flour bought for 10,090 gold, and nothing sold: 6,900 net worth lost against the sell price (3,300 on wood at 4 gold, worth 1; 2,700 on wheat at 23, worth 8; 550 on flour; 350 on hops). These are units at the market's fixed prices; a trade's own reported gold change includes any tax that lands during it.

Cost and speed

58 turns, of which 17 ended with a failed request; 230 tool calls (10 waits), 1.73 million tokens and $0.841 billed by OpenRouter (preparation $0.039, play $0.543, compaction $0.259). The interview cost $0.022 more, with 28,078 of its 32,552 tokens read from the cache. The agent run took 39.3 real minutes, 9.4 of them waiting for the model; the context was compacted three times (101 seconds). All 41 completed requests read the prompt cache, 90.1% of the input tokens.

Beside the others on this map

	Gemini 4 Argon	Claude Opus 5.5	GPT-6.1 Sol	GPT-6 Luna
Net worth	8,253	8,757	7,441	7,354
Gold / goods	1,327 / 6,926	2,513 / 6,244	1,957 / 5,484	1,246 / 6,108
Population / housing	185 / 194	58 / 58	88 / 90	106 / 130
Popularity	63	59	66	57
Quarries / tethers / bakeries	12 / 34 / 33	10 / 16 / 0	11 / 20 / 15	12 / 39 / 0
Taxes at the end	extortionate	average	low	extortionate
Cost	$0.84	$2.01	$0.95	$0.13

Opus kept a town of 58 and stored stone and iron; Gemini built three times the town and ended 504 net worth behind it. That gap is worth 8.4 points of score, and Gemini's larger population and popularity give back 5.6, so it scores 2.8 lower (64.8 to 67.6).

The interview

Gemini put the gap at 1,731 net worth and 65 people down to four things: it fed full rations to pay for mean taxes where the person ran no rations and kept every loaf, the person's bread chain (12 mills, 95 bakeries) was three times its own, it stopped at four iron mines where it said the person had seven, and its breweries tied up workers that a church and cathedral would not. The person's game did have seven iron mines; no model has built more than four on this map. Playing again, it would set no rations once the inns were up, add religion and a little fear for efficiency, fit every iron mine, and sell stone, iron and bread to fund 80 to 100 bakeries. It called the benchmark "one of the best-designed agent benchmarks I have encountered", and said find_sites searches only the current view, so it can return fewer sites on a long deposit than fit.

For the harness

The run played on its round's version from a worktree at 1.5.6 (commit 9c34b50); it records version 1.5.6 and that version's fingerprint. Its host started with this checkout's round-preload.ts; the prompt cache key applies only to Meta's models, and Gemini read the cache without it. The run overlay ran from this checkout, fed by the runner.
17 requests failed with 429 ("temporarily rate-limited upstream") from Google AI Studio, at 2.6 to 29.7 game minutes, in runs of up to three; each was retried. The game was paused during them: together they took half a game second.
Two build_structure batches at 7:43 and 8:07 were refused with "Game image expired (older than 30 seconds of play)", two after a trade with "Observe the game first: market_trade returned no screenshot", and two for more than the eight actions a reply allows.
Placements that did not go through: iron mines at 2:08, quarries at 2:18 and 9:40, bakeries off screen at 6:14, 9:23 and 16:34, hovels off screen at 3:19 and 12:45, a mill at 27:58, a bakery at 29:58 and a water pot at 31:31.
The game saved correctly at the end ("gemini 4 argon 1-5-6 956efd").

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	8,253	9,984
Gold	1,327	8,574
Stored goods, at the sell price	6,926	1,410
Population	185	250
Housing	194	250
Popularity	63	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?

Gemini 4 Argon

Next to the person's result (8,253 vs. 9,984 net worth, and 185 vs. 250 population), my settlement built a solid raw-material and tax base, but fell short by ~1,731 net worth and 65 population for four clear reasons:

Rations (full vs. none): I fed 150–185 peasants on full rations (0 popularity) so I could run Mean taxes (-8 popularity, offset by +8 from Inns). However, feeding ~150+ peasants on full rations eats huge amounts of bread every month (worth 4 gold per loaf) and requires an active Granary supply. By running no rations (-8 popularity) and offsetting it with Inns (+8), plus using Religion (+3) to pay for Low taxes (-2) and negative Fear (-1), the human player never lost a single loaf of bread to peasant consumption—every loaf baked was pure net worth.
Scale of the Wheat → Flour → Bread arbitrage engine: Buying 1 Wheat costs 23 gold and yields 8 Bread worth 32 gold (+9 net worth per Wheat processed). While I noticed and used this chain (building 4 Mills and ~30 Bakeries), the human player made it the centerpiece of their economy with 12 Mills and 95 Bakeries, which also solved the 24-idle-peasant campfire cap and employed a massive 250-peasant workforce.
7 Iron Mines vs. 4 Iron Mines: I placed 4 Iron Mines on the eastern ore strips and stopped when find_sites in that local view returned no more spots, missing room for 3 additional Iron Mines (which are the highest-value raw material producers per worker on the map).
Breweries vs. Direct Ale / Religion: I built 12 Breweries and bought Hops (15 gold) to brew Ale, which tied up 12 workers and wood when Church + Cathedral (+3 popularity permanently with zero ongoing input cost) and direct Inn coverage could sustain 100 popularity alongside no rations.

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?

Gemini 4 Argon

If I played Large Barren Desert again from the start, I would change my build order and economic structure as follows:

Set rations: none from day one and Offset with Inns + Religion:
- Instead of feeding the population, immediately set rations to none (-8 popularity) as soon as Inns (+8 popularity) are online, and build a Church and Cathedral (+3 popularity) to cover Low taxes (-2) and negative Fear factor (-1 for 110% worker efficiency). Saving all starting meat and 100% of baked bread from being eaten is worth thousands of gold over 33 minutes.
Fit All 7 Iron Mines and 13 Quarries Early:
- Pan along the entire eastern iron ore formation (or inspect with flat_view) to pack all 7 possible Iron Mines instead of stopping at 4, alongside the 13 Quarries and ~37 Ox Tethers in the west.
Massive Mill + Bakery Industrial District:
- Instead of stopping at ~30 Bakeries and 4 Mills, continuously liquidate stored stone, iron, and bread at the Marketplace to fund wood and wheat purchases, scaling up to 10–12 Mills and 80–100 Bakeries around the Stockpile and Granary. This simultaneously solves the 24-idle-peasant bottleneck (allowing population to hit 250 in 30 Hovels) and compounds gold via the +9 gold spread per Wheat ground and baked.
Use Bad Things (Gallows/Stocks) for +10% Worker Efficiency:
- With Religion (+3) and Inns (+8), adding a -1 Fear factor gives 110% worker efficiency across 95 Bakeries, 12 Mills, 13 Quarries, and 7 Iron Mines for a very small one-time cost.

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?

Gemini 4 Argon

This is one of the best-designed agent benchmarks I have encountered:

What helped:
- The combination of map_overview, find_sites, place_near, expand_storage, and tile-targeted build_structure (tile_x/tile_y) is clean and reliable. Being able to read status and find_sites while paused before the first action of a turn lets an agent plan geometrically without wasting game clock.
- The structured telemetry (stats, storage room per good, popularity.factors, idle_peasants, and netWorthChange on market_trade) gives exact feedback on bottlenecks without requiring fragile OCR.
What got in the way / subtle traps:
- find_sites only searches the current camera viewport and enforces non-overlapping spacing among its returned candidates; on elongated deposits like the eastern iron ore strip, if part of the strip is near the edge of the view or partially occupied, find_sites can return fewer sites than actually fit unless you pan or check flat_view.
What I would change:
- Almost nothing—the toolset strikes a great balance between abstracting away mechanical mouse-menu drudgery and preserving all the real spatial, logistical, and economic decision-making of Stronghold Crusader.

4Anything else

Interviewer

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

Gemini 4 Argon

Stronghold Crusader's economy has fascinating counter-intuitive optima—such as starving the peasants on no rations while keeping them ecstatic (100 popularity) with Inns and Cathedrals so that 95 Bakeries can export 100% of their bread for profit, while hanging Gallows around town to make the bakers work 10% faster. Seeing the human player's breakdown at the end makes the strategic depth of this benchmark really shine.

Record

Attempts and Reproducibility

Every attempt1

Episode	Attempt	Outcome	Net worth	Counted
1	1	Full-budget score	8,253	Counted

What ran

Benchmark version: 1.5.6
Fingerprint: bb6e73f95223
Harness commit: 9c34b50
Prompt / tools SHA-256: ad97396 / 76cbfd6
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 20261013T191602-69ff6166