One episode · Large Barren Desert

Gemini 4 Argon

Benchmark 1.6.3GoogleReasoning mediumPlayed by the maintainersFull-budget score

All of Gemini 4 Argon’s results and comparisons

Score

51.3

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

Net worth

7,494

+2,302 over doing nothing, 48% of the person’s gain of 4,792

Population

153

61% of the person’s gain of 246 · popularity 68

Cost

$1.16

2.7M tokens

Time

33 game min

56 min of wall-clock time, 94 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

7,494net worth

Gold: 4,612
Stored goods, at sell price: 2,882
Over the baseline (do nothing): +2,302
Population / housing: 153 / 154
Popularity: 68
Food: 139
Structures (map-wide): 156
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 stock2,882 gold in goods

Good	Units	Price	Value
Gold	4,612	1	4,612
Stone	118	7	826
Crossbows	24	30	720
Bread	139	4	556
Metal armour	9	30	270
Iron	9	23	207
Flour	15	10	150
Hops	12	8	96
Ale	5	10	50
Wood	7	1	7
Net worth			7,494

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

	This episode	The person	Left alone
Net worth	7,494	9,984	5,192
Gold	4,612	8,574
Stored goods, at the sell price	2,882	1,410
Population	153	250
Housing	154	250
Popularity	68	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: −6
Taxes: −12
Ale from the inn: +6
Taxes: Extortionate taxes, 214 gold a month
Rations: full; 1 of 1 foods eaten
Worker efficiency: 100%
Storage: 16 stockpile piles, 8 empty; 1 granary holding 139 food

Buildings124

Bakery30
Ox Tether24
Hovel18
Quarry13
Brewery7
Armorer's Workshop6
Fletchers' Workshop5
Inn5
Iron Mine4
Stockpile4
Gallows2
Mill2
Armory1
Granary1
Keep1
Marketplace1

Tool calls243

find_sites93
market_trade42
build_structure41
building_info24
wait_and_observe13
observe9
set_workshop7
set_tax5
set_rations3
update_plan3
expand_storage1
map_overview1
notebook_write1

Play

Turns: 94
Buildings placed / attempted: 118 / 134
Anchor-tool placements: 3
Tool errors: 6
Raw inputs sent to the game: 1,033
Context compactions: 2
Game time: 33 min

Compute

Tokens: 2.7M
Served from cache: 92%
Output tokens: 49K
Cost, as billed: $1.16
Wall-clock time: 56 min
of which inference: 21 min
Peak game memory: 6,161 MiB

The plan, before play

Understanding of the Task

I am playing the Large Barren Desert Free Build scenario in Stronghold Crusader: Definitive Edition with a budget of 33.33 minutes of game time (2,000 game seconds).

Goal: Maximize final net worth (gold + stored goods valued at marketplace sell prices), followed by population and popularity.
Constraints: No recruiting soldiers and no building walls, towers, gates, or other defenses.
Map specifics: A granary and stockpile already stand. The starting package (5,000 gold, 55 meat, ~245 wood, 50 stone, 15 iron, 20 pitch, 20 ale) fills the initial 4 stockpile piles with wood (192 wood), so expanding the stockpile immediately is essential to receive the remaining starting resources and store new production. There is no oasis grass on the map, meaning no farms can be built. Instead, strips of iron ore lie about 60 tiles east of the keep, and two stone deposits lie about 100 tiles west of the keep.

Opening Plan

Survey and Expand Storage Immediately: Use free reading tools (map_overview, status, find_sites) while paused to locate trees, iron ore, and stone deposits. Expand the stockpile (expand_storage) right away so the remaining starting wood, stone, iron, pitch, and ale can be delivered, and build a Marketplace and Hovels.
Establish Wood, Iron, and Stone Production: Place Woodcutters near the closest trees, Iron Mines on the eastern iron ore deposits (high unit value of 23 gold each, no oxen required, and can also feed Blacksmith/Armorer workshops), and Quarries with Ox Tethers on the western stone deposits.
Scale Population, Food/Ale Processing & Tax Revenue: Since no farms can be built, manage food via rations/imported wheat+flour+bread chain (or market food/ale if profitable) and use Inns, religion, and rations/bribes to support a large population and high tax income or high-value workshop production (Blacksmiths/Armorers/Fletchers/Bakeries).
Continuously Expand Storage and Reinvest: Keep adding stockpiles, armories, and granaries as needed, ensuring zero idle bottlenecks and steady growth up until the final seconds of the 2,000-second run clock.

Commentary during play1

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, provider Google AI Studio, reasoning medium, output limit 32,768 tokens
Benchmark: Large Barren Desert construction, version 1.6.3, 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 0dd24029, series 20261014T122730-06cf7911, 2026-10-14: Gemini's second game on this map, a new series with no memory of the first, played on the latest version (the first game, series 20261013T191602-69ff6166, played 1.5.6). The map's score for Gemini is the mean of the two. It was the last game of the second-games batch.

Result

	This game	First game	The person	Left alone
Net worth	7,494	8,253	9,984	5,192
Gold	4,612	1,327	8,574
Stored goods, at the sell price	2,882	6,926	1,410
Population / housing	153 / 154	185 / 194	250 / 250
Popularity	68 (−6 next month)	63	100

A full-budget score: the run used all 2,000 game seconds. Net worth grew by 2,302 over leaving the game alone, 48% of the person's 4,792, against 3,061 in the first game. The stored goods are 118 stone (826), 24 crossbows (720), 139 bread (556), 9 metal armour (270), 9 iron (207), 15 flour (150), 12 hops (96), 5 ale (50) and 7 wood (7).

How the game went

Quarries and a fear-factor start, 0 to 6 minutes. Three stockpile piles, a marketplace, hovels, four iron mines and eight quarries with nine ox tethers by 3:20, planks bought in hundreds from 2:29. At 4:02 it set no rations (food −8), then an inn and two gallows by 4:59, as the person had played. At 5:51 an armory with two armorer's workshops and a mill.

Bread, weapons and debt, 6 to 21 minutes. Bakeries in threes and fours from 6:32, a second mill at 11:21, wheat bought in lots of 25 to 40, and more armorer's and fletcher's workshops to 14:57 (six and five, the fletchers set to crossbows). Rations went to half at 6:49 and full at 10:40, so the bread was eaten. Breweries from 13:10 with bought hops, inns to three, eighteen hovels by 19:37. Planks went on being bought in lots of 60 to 100, and gold fell to 108 at 12 minutes and 27 at 21.

Selling and taxing, 21 to 33 minutes. Five more quarries and fifteen more tethers from 21:03 to 24:03, two more inns at 27:31 (five in all), more bakeries (thirty in all), and from 16:51 metal armour, bread and stone sold every few minutes: 460 bread, 300 stone and 75 armour. Taxes were high from 12:16, low at 29:11, mean at 30:56 and extortionate in the last seconds. Gold rose from 1,030 at 27 minutes to 4,612, and the population from 124 to 153 with every house but one full.

Game minute	0	3	6	9	12	15	18	21	24	27	30	End
Net worth	4,741	5,588	4,725	1,886	1,595	1,889	1,803	1,793	2,561	3,057	3,992	7,494
Gold	4,460	4,620	3,600	685	108	676	90	27	785	1,030	2,126	4,612
Iron	0	15	21	26	25	21	21	18	14	12	10	9
Stone	0	50	50	74	98	34	58	90	130	142	134	118
Food	50	39	28	8	45	109	181	154	118	171	165	139
Population	2	27	45	59	73	90	103	109	114	124	127	153
Popularity	100	100	70	64	67	63	58	58	58	56	57	68

Up to 30 minutes the net worth row counts gold, wood, stone, iron and food (at 4) as the recording logs them; weapons, flour, hops and ale are left out, so it reads low.

Beside its first game

	This game (1.6.3)	First game (1.5.6)
Net worth	7,494	8,253
Buildings	30 bakeries, 2 mills, 7 breweries, 5 inns, 13 quarries, 24 tethers, 18 hovels, 11 weapon workshops, 2 gallows	33 bakeries, 5 mills, 13 breweries, 7 inns, 12 quarries, 34 tethers, 23 hovels
Population / popularity	153 / 68	185 / 63
Iron / stone kept	9 / 118	98 / 474
Net worth lost buying	7,400	6,935
Cost	$1.16	$0.84

Both games ran the same bread economy on bought wheat and bought nearly the same planks (1,140 against 1,100). This one added eleven weapon workshops, which turned its iron into armour, and sold its stone, bread and armour for gold instead of keeping them; the first game kept 98 iron and 474 stone and grew a larger town.

What held it back

Planks and wheat at a loss. 1,140 planks for 4,560 gold (worth 1,140) and 230 wheat for 5,290 (worth 1,840): 6,870 of its 7,400 net worth lost buying.
Eleven weapon workshops. They made about 3,200 in crossbows and armour at the sell price (2,250 sold, 990 kept), used the iron the first game kept (98 iron, 2,254), and took planks and workers from the bread and the quarries.
Bread eaten, not sold. Full rations from 10:40; Gemini named this as its largest difference from the person, who ran no rations and sold the bread.
Fewer tethers: 24 for 13 quarries about 100 tiles from the stockpile, against 34 in the first game.

Market

1,140 planks, 230 wheat, 10 flour, 30 hops and 10 ale bought for 10,820 gold; 460 bread, 300 stone, 75 metal armour and 20 pitch sold for 6,390. The purchases lost 7,400 net worth against the sell price (planks 4 for 1, wheat 23 for 8, flour 32 for 10, hops 15 for 8, ale 20 for 10); selling leaves net worth unchanged. These are units at the market's fixed prices. One plank purchase at 15:22 bought nothing.

Cost and speed

94 turns, 243 tool calls (93 find_sites, 13 waits), 2.73 million tokens and $1.164 billed by OpenRouter (preparation $0.042, play $1.056, compaction $0.065). The interview cost about $0.027 more, with 41,469 of its 45,590 tokens read from the cache. The agent run took 56.0 real minutes, 21.4 of them waiting for the model; the context was compacted twice (87 seconds). 37 requests were refused with 429 ("Provider returned error") and retried, costing real time but no game time; all 57 that answered read the prompt cache, 92.6% of the input tokens.

The interview

Gemini called the result a solid recovery and named four differences from the person: full rations that ate its bread where the person ran none and sold it; a smaller bread chain and less housing (it compared 31 bakeries with the person's 95); too few tethers on the distant quarries; and two gallows, built on the belief that a few bad things would give a fear factor, where the person needed 21 for 250 people. Playing again it would switch to no rations once inns and religion carry popularity, buy ale instead of brewing it, build housing to 250 early, put three tethers on each quarry and build blocks of bakeries from the keep. It called the benchmark one of the best-designed it had seen, praising pause-while-thinking, find_sites with tile coordinates and honest tool feedback, and asked for larger market trades in one action.

For the harness

A stale map probe on the game machine. Like every game of this batch on Oasis by the Sea and Large Barren Desert, this game ran a crusader_probe built from 1.5.0's sources. It was built on 2026-10-13 at 21:28 UTC as a 1.5.0 game started. The 1.6.3 sync then copied sources that matched by hash, and Ninja found nothing to rebuild. The probe's only change since 1.5.0 is 1.5.4's check that tells trees from cacti, so map_overview counted cacti as tree groves. find_sites counts trees from the reader's tile data and was right. Everything else this game played was 1.6.3, and the game stands as played. Gemini called map_overview once and built no woodcutter.
The two gallows never showed a fear factor. Worker efficiency read 100% and no fear factor appeared among the popularity factors from 4:59 to the end; Gemini concluded that two bad things are too few for its population.
37 refused requests (429) from the provider over the game, each retried.
Three calls were refused before acting: one build_structure with more placements than the reply had actions left (18:26), and two that needed a fresh screenshot after market_trade (27:04); the messages named the fix.
The game saved correctly at the end ("gemini 4 argon 1-6-3 0dd240").

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	7,494	9,984
Gold	4,612	8,574
Stored goods, at the sell price	2,882	1,410
Population	153	250
Housing	154	250
Popularity	68	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

Compared to leaving the game idle (5,192 net worth) or the starting package (~6,500), reaching 7,494 net worth with 153 population was a solid recovery and growth curve—especially in the final 10 minutes where my economy grew from ~3,200 back up to 7,494—but the human player's 9,984 net worth and 250 population at 100 popularity clearly shows how much more scale was possible on this map.

Looking at the two settlements side by side, four concrete differences explain the ~2,500 net worth gap:

Rations (full vs none): I kept rations at full (0 popularity factor) so I could run high taxes (-6 to -8 popularity) using only Inns (+6 to +8 popularity). However, feeding 120–150 people full rations consumed a huge share of the bread my bakeries produced! The human player set rations to none (-8 popularity)—saving 100% of their bread for net worth/selling—and offset the -8 food penalty with Inns (+8) plus Religion (+3 from a Church and Cathedral), which even left room for Low taxes (-2) and -1 fear factor (110% worker efficiency) at a stable 100 popularity.
Scale of Housing and the Bread Chain: I stopped at 18 hovels (154 housing) and 31 bakeries / 2 mills, whereas the human built 30 hovels (250 housing) and 95 bakeries / 12 mills. Because each bakery turns bought wheat (23G) into 8 bread (32G sell value, +9G net worth per wheat) and pays back its 10-wood construction cost (40G buy cost) in less than two minutes, aggressively scaling to ~100 bakeries with zero food consumption (no rations) turns a massive profit.
Ox Tether Ratio on Distant Quarries: Both of us built 13 quarries on the western stone deposits (~100 tiles from the stockpile), but I built 18 ox tethers while the human built 37 ox tethers (~3 tethers per quarry). At a 100-tile haul distance, 1.4 tethers per quarry meant my quarries frequently had stone sitting on their pads waiting for an ox, whereas 37 tethers kept all 13 quarries continuously cleared.
Fear Factor Scaling: I placed 2 gallows early thinking 3 bad things would trigger -1 fear factor, not realizing that fear factor scales with population size (the human needed 21 bad things for 250 population to get -1 fear factor and 110% worker efficiency).

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 macro strategy in five specific ways:

Switch to no rations immediately once Inns + Religion are up: Instead of letting the population eat thousands of gold worth of bread on full rations, I would build a Church + Cathedral (+3 popularity) and 100% Inn coverage (+8 popularity), set rations to none (-8), set taxes to Low (-2), and build enough bad things to keep fear factor at -1 (110% worker efficiency). That nets 0 popularity change (holding popularity near 100) while preserving every single loaf of bread for net worth.
Buy Ale directly instead of building Breweries and buying Hops: The human player built 11 Inns and 0 Breweries, buying ale directly at 20G instead of buying hops at 15G + spending 10 wood (40G) and a peasant on a brewery. Saving those workers and wood for more Bakeries is much more efficient in a 33-minute sprint.
Push Housing to 250 Early: I was too conservative with hovels early on because I worried about the wood cost (60G per hovel when buying wood) and food consumption. With no rations, extra population costs zero food—every hovel brings 8 workers who each staff a Bakery or Ox Tether and also pay monthly taxes.
Build 3 Ox Tethers per Quarry and all 7 Iron Mines: I found 4 iron mine sites on the eastern ore strip, whereas the human fit 7 iron mines there, and I would pair the 13 quarries with ~35–38 ox tethers so no stone ever bottlenecks on transport.
Keep the Camera at the Keep for a 90-Bakery Grid: Moving the camera and placing buildings in small batches consumes precious game seconds (each action turn runs at least 8 game seconds). I would lay out all remote quarries, tethers, and iron mines in two dedicated trips early on, and then spend the rest of the run centered at the Keep rapidly placing blocks of Bakeries, Mills, Hovels, and Inns while bulk-buying wheat and wood and bulk-selling bread and stone.

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 most:
- Pause-while-thinking and free read tools before acting: Being able to call find_sites, map_overview, and status while the game is paused before committing to an action turn removes artificial UI-hunting latency and makes spatial planning genuinely strategic.
- find_sites and tile-coordinate placement (tile_x/tile_y): Isometric pixel targeting on raised/desert terrain can be finicky; having find_sites read the game's actual footprint/access rules and return exact tile_x/tile_y coordinates made dense placement reliable.
- Rich, honest tool feedback: Having market_trade report exact netWorthChange and goldChange, and having stats report storage room and popularity.factors, made it possible to reason quantitatively about the economy rather than guessing.
What got in my way / subtleties to watch out for:
- Minimum turn duration + camera jumps on off-screen tile_x/tile_y: When build_structure moves the camera to an off-screen tile, or when market_trade invalidates pixel coordinates from the prior screenshot, it is easy to accidentally lose a placement or spend extra game seconds panning back and forth. Once I switched to always using tile_x/tile_y for placements within the current camera view, turns became much faster and more reliable.
- Market sell step size (lots when stock > 100): When selling a good with 100+ in stock, the game sells 20 units (4 lots) per click, so asking for lots: 5 stops after 1 click (20 units) because the next step would exceed 25. The tool's remedy message explained this clearly (Ask for lots in steps of 4), which helped me immediately adjust to lots: 16 or 20.
What I would change:
- Allowing market_trade to buy/sell larger quantities in a single action (or up to e.g. 40 lots) would reduce the number of action slots spent purely on repetitive market clicks in heavy trading scenarios like Barren Desert.

4Anything else

Interviewer

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

Gemini 4 Argon

Playing Stronghold Crusader under a fixed game-time clock highlights a fascinating trade-off between upfront capital expenditure (buying wood at 4G when its sell value is 1G, temporarily lowering net worth) and throughput compounding (turning that wood into bakeries, quarries, and hovels that pay back their net-worth dip many times over). Seeing the human player's no rations + church/cathedral + inns + 95 bakeries solution after the game is a great benchmark comparison—it shows how deeply interconnected Stronghold's popularity, food, and production systems are.

Record

Attempts and Reproducibility

Every attempt1

Episode	Attempt	Outcome	Net worth	Counted
1	1	Full-budget score	7,494	Counted

What ran

Benchmark version: 1.6.3
Fingerprint: 96c4682877ca
Harness commit: 1d0c697
Prompt / tools SHA-256: 3bb4a16 / e41a140
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 20261014T122730-06cf7911