Gemini 4 Argon

Written by Claude (drafted by Sonnet 5.5, checked by 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: Oasis by the Sea construction, version 1.7.3 (commit b967c65), 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. Weapons and armour in the armory count toward net worth at the rules' prices, as since 1.6.0.
• Run ffe69c40, series 20261017T132016-ab6eb16c, 2026-10-17: Gemini's weapons-counted game of the 1.7.3 batch, a fresh series with no memory of any other game. Its no-arms game failed on an OpenRouter key limit and will be played again; it is not in the tables.

Result

A full-budget score: the run used all 2,000 game seconds. Net worth grew by 11,954 over leaving the game alone, 65% of the person's 18,293. With 167 people and popularity 69 the score is 67.4 by the project website's formula (80% net worth, 10% each population and popularity, the net-worth share clipped to −1..2); this game counts on the website's leaderboard. It is 0.6 below its 1.6.3 game and 16.4 above its 1.5.1 game; with weapons counted for nothing it would be 61.8.

The stored goods are 286 iron (6,578), 358 stone (2,506), 31 crossbows (930) and 19 pikes (342), 102 wood (102), 20 bread (80), 5 flour (50) and 4 wheat (32). Iron and stone are 69% of net worth, weapons 9.7%. The town ended with 23 idle peasants, 19 empty places in its hovels, 13 of 32 stockpile piles empty and 20 food, all bread.

How the game went

Times are the game clock when a call began (the recording's clock; the run clock in the observations reads a few seconds later). The game is paused while the model replies; a build call keeps it running for 13 to 67 seconds (median 32) while it clicks.

The opening, 0 to 5 minutes. One reply of reading and planning, then at 0:00 the marketplace and two stockpile piles, 100 planks at 0:13, and in the calls from 0:28 to 4:03 all twelve iron mines, hovels, ten woodcutters (by 4:42), two quarries and orchards. Planks were bought four more times; gold was 0 at 3:07 and the first iron came at 2:43. At 3:07 it set taxes to mean (34 gold a month, −8 popularity a month); popularity fell from 100 to 78 by 4:34 and it set low taxes at 4:35.

Food, stone and the first workshops, 5 to 16 minutes. Three quarries and three ox tethers at 5:47; at 7:57 a wheat farm, mill, granary, the first armory and two fletchers (after an empty reply that cost 8 game seconds); at 9:14 three bakeries, two fletchers and a poleturner; cheese, meat and ale bought at 11:42 to 11:51 and an inn at 11:55; at 12:49 two bakeries and a poleturner. Taxes were at none from 11:35 to 12:43. Popularity fell from 74 at 5:32 to 53 at 11:35. Gold stayed under 500 and it sold 45 iron in five lots to pay for building. Population rose from 42 to 88.

Weapons and bread, 16 to 25 minutes. Workshops at 16:41, 18:03, 19:20, 20:10, 21:40 and 23:50 (see Weapons), bakeries (14 stand), a second mill and two more breweries; average taxes from 19:17, extra rations from 23:48. Idle peasants rose from 1 at 16:20 to 19 at 23:48.

The last 8 minutes. No more workshops. A 25:07 call placed three of four bakeries and neither of two hovels for want of wood (35 at its start, 0 by 25:44); 65 planks were bought at 26:13. Then an inn, a wheat farm, a mill and hovels (26:28 to 29:19) and seventeen woodcutters (26:28 to 32:38), which its reasoning from 29:15 to 32:35 gave as jobs for idle peasants. A popularity warning at 27:54 (59 to 48, under 50 for 33 s) was answered in the next call with low taxes and extra rations; a food warning at 30:48 (68 food for 153 people on extra rations) with full rations. Taxes went to high at 32:35.

Each column is the last reading at or before that minute. Up to 30 minutes the net worth row counts gold, wood, stone, iron and food (at 4) as the recording logs them; weapons, flour, wheat, hops and ale are left out, so it reads low (the weapons, 864 at 29:16 and 1,272 at the end, are in the End value only).

Weapons

It built two armories, fifteen fletcher's and seven poleturner's workshops, and no blacksmith, armorer or tanner. Its opening plan ended with "Weapon Workshops + Armories for maximum net worth". At 7:48 its reasoning priced a crossbow at +27 over its wood and a pike at +16, with wood in surplus (31.7 a minute at 7:49), and the first armory and two fletchers went up in the call that began at 7:57. Fletchers followed at 9:14, 14:49, 16:41, 18:03, 19:20, 20:10 and 21:40 (15 in all), poleturners at 9:14, 12:49, 16:41, 18:03, 21:40 and 23:50 (7), the second armory at 14:49. Each was set to crossbows or pikes by set_workshop (22 calls, 157 game seconds; they start on bows and spears). At 25:02 it wrote that more workshops would return only a fraction of their cost in the time left, and built no more.

Output came slowly. The workshops waited for staff (12 open jobs at 10:07, 2 at 12:41) and a fletcher inspected at 13:46 was working, but the armory read 0 stored at 14:48 and 16:20. At 14:49 the model saw that nothing had been delivered; at 16:20 it traced this to a bakery it had placed at (420, 361) in the 12:49 call, which in its reading closed the corridor between the marketplace and the keep and cut off both armories (the second had to touch the first). It demolished the bakery at 16:20 and the armory held 4 weapons at 17:32. At the rate measured later, only a few weapons were due between 13:46 and 16:20, so the logs do not show what the block cost. At 18:41 it suspected that hovels and an inn walled in the five hill workshops at (403 to 410, 351 to 362); two demolish calls (18:41, 18:49) hit buildings drawn in front and it did not try again. Whether those five delivered is not in the logs.

The stock was 96 in weapons at 17:32, 330 at 22:24, 660 at 26:13, 1,194 at 32:35 and 1,272 at the end (31 crossbows, 19 pikes), 48% of it after 26:13. From 22:24 the 15 fletchers made 23 crossbows (0.14 a workshop a minute, about 4 gold) and the five to seven poleturners 14 pikes (0.2, about 4 gold); in the same readouts an iron mine delivered 0.9 to 1.0 iron a minute, 21 to 23 gold. None was sold or bought; the armories held 50 of 100.

From prompt/game-mechanics.md's table the workshops cost 2,200 gold (22 × 100) and 370 wood (15 × 20, 7 × 10), the armories 10 wood, and the weapons used 131 wood (31 × 3, 19 × 2) and no iron. At store value that is 2,200 gold and 511 wood against 1,272: 1,439 behind (6.3 points), or 928 behind (4.1 points) against the gold alone. Weapons added 5.6 points to the 67.4. Gold was short while they went up: 1,800 was committed by the 20:10 call, gold in hand had not reached 500 since 3:00, and it sold iron or stone at 9:09, 12:45 and 17:32 with the workshops as its stated reason. Searches for more mine (4:35, 23:48, 25:02) and quarry sites (6:38, 23:48) found none. Held, the gold would count 1:1 (2,200 is 9.6 points); a chapel (250 gold, looked up at 4:35, 10:07 and 22:24, never built although "Religion" was in the opening plan) is the other use its logs show it pricing. The workshops also gave 22 jobs: at 23:48 it placed the last two poleturners to employ 19 idle peasants.

Beside its other games and the other models

Against its 1.6.3 game this one ended 429 lower: gold −1,876 and bread −1,900, against stone +1,449, weapons +1,272 and iron +598. The score moved 0.6: net worth −1.9, population −0.2, popularity +1.5. For this model, then, weapons do not explain the move between versions. The 1.6.3 game built none and scored 68.0; this one built 22 workshops, and without their output it would stand 6.2 below that game, because the rest of the town was 1,701 lower (half the bakeries, a quarter of the mills, 20 bread against 495). The workshops' 2,200 gold is of the order of the 1,876 gold shortfall.

Among the eight finished weapons-counted games of the batch (Muse Wonder's are to come) 67.4 is second, behind GPT-6.1 Sol's 82.3 (16,661, with 195 in weapons from five fletchers) and ahead of Fable 37.5, Sonnet 36.5, Opus 32.9, Luna 25.7, Haiku 23.0 and GLM 18.9. Its 1,272 in weapons is the most of the eight (Haiku's 810 and GLM's 548 next) and its 167 people the most (Sol 150). Among the 1.6.3 round's second games it would stand second to its own 68.0, ahead of Sol's 54.6; the best of the 1.5.x round on this map was 52.0 (Opus).

What held it back

• Weapons. 1,272 came back against 2,200 gold and 511 wood, and the first 96 only at 17:32.
• Food ran down. Food fell from 170 at 23:48 to 20, with rations extra for 5½ minutes of it (23:48 to 26:26, 27:56 to 30:48). At 23:48 its reasoning counted a surplus (47 food a minute against 46 eaten); food fell 62 in the next 2 min 25 s. Bakery output fell from 47 a minute at 29:16 to 16 at 32:36 while the mills delivered 1.3 to 5 flour a minute. The 150 food lost was worth 600 (2.6 points).
• Little farmland used. Three wheat farms and no hops farm. Farm searches found a spot (usually one) at 7:57, 10:07, 12:43, 13:40 and 25:02 and none at 9:09, 14:49, 16:20, 17:32, 27:56 and 29:15; two hops farm placements (6:38, 14:49) found no room. The breweries ran on 10 bought hops.
• Taxes against popularity. Mean taxes at 3:07 took 22 points in about 90 s for 34 gold a month; popularity was 53 at 11:35 and 48 at 27:50.

What worked

• Twelve iron mines by 4:03, which dug 336 iron (286 stored, 50 sold): half of net worth in store, at about 0.9 iron a minute each.
• Stone. Five quarries and 13 tethers gave 433 stone (358 stored, 75 sold), 13 to 27 a minute in the readouts from 19:59.
• People. 167 of 186 places, the most of the batch, with popularity 69 and the inns at +4.
• Placements and warnings. Every placement reported as placed is in the final count (15 bakeries placed, one demolished; the stockpile and granary the save starts with included); 148 of 168 requested went through (no room 13, wood 3, rejected 2, one unverified, one off screen). The warnings of 27:54 and 30:48 were answered in the next call, and the one refused call (a sale of 15 stone at 20:07, when the game sells 20 a click) was redone as 40 at 21:36, as its remedy said.

Market

460 planks (4 each, worth 1), 5 flour, 10 cheese, 10 meat, 5 ale and 10 hops were bought for 2,410 gold; 50 iron and 75 stone sold for 1,675. The purchases lost 1,690 net worth against the sell price (7.4 points; planks 1,380); selling leaves net worth unchanged. 395 planks were bought by 5:36 for the opening and 65 at 26:13. These are units at the market's fixed prices. The 23 market calls took 148 game seconds.

Cost and speed

37 turns, 282 tool calls (136 find_sites, 30 build_structure, 23 market_trade, 23 waits, 22 set_workshop, 20 building_info, 8 set_tax, 6 observe, 4 set_rations, 3 update_plan, 3 demolish_building, 2 map_overview, 2 inspect_building), 2.00 million tokens and $1.700 billed by OpenRouter, every request to Google AI Studio (preparation $0.044, play $1.489, compaction $0.167). The agent run took 54.7 real minutes, 23.7 of them waiting for the model (38 s a reply, longest 86 s) and 2.3 more on four compactions (6:37, 13:39, 19:58, 27:54). Fast-forwarding the 22 waits saved 205 s of real time. 35 of the 37 gameplay requests read the prompt cache, 82.2% of the input tokens; the two that did not were at 7:48 and 27:56 (the first request after the last compaction), and no gap between a reply and the next request passed 2 minutes.

Of the 2,000 game seconds, the calls other than waits took 1,394: build_structure 1,014 (30 calls, 148 buildings placed, 6.9 s a building, 4.9 a call), set_workshop 157, market_trade 148 and the rest 75; the waits took 585. Its 1.6.3 game came to 8.7 s a building (49 calls, 144 buildings) and its 1.5.1 game to 6.5 s (58 calls, 141 buildings). Neither ran with the host on the game machine; this one did.

The interview

It did not run. Its request, made after the game was saved, was refused at once with a 402 from OpenRouter (the key's spending limit: it could pay for 30,386 output tokens and 32,768 were asked), so there are no answers to check and no interview cost.

For the harness

• The granary line. The rules' Start paragraph says to place the required granary; the save has one (docs/benchmark.md). No call tried: the first reasoning saw it built. A second granary placed at 7:57 was accepted.
• A bakery and the armory. The bakery at (420, 361) was accepted at 12:49 with no access note, though docs/tools.md (Routes to the keep) says a footprint that closes another building's last way reports cuts_route. The model read it as cutting off the armories; the logs do not show whether the tile map saw that corridor as open, or whether the block, rather than slow workshops, kept the armory empty until 16:20.
• An empty reply at 7:48 (6 output tokens) made the host let 8 game seconds pass; its note says four in a row end the run.
• Demolition refused as covered at 18:41 and 18:49; the remedy (aim at the back tile, or demolish the front building first) was not tried.
• Workshop products. Every fletcher and poleturner starts on the cheaper product, so each of the 22 needed a set_workshop call (about 7 game seconds each, 7.9% of the budget).
• The interview failed on the key limit and ended there; no gameplay request failed.
• The game saved correctly at the end ("gemini 4 argon 1-7-3 ffe69c").