GPT-6.1 Terra

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: GPT-6.1 Terra (openai/gpt-6.1-terra) on OpenRouter, served by OpenAI or Azure (no other fallbacks; all 99 billed requests were billed to OpenAI), reasoning medium, output limit 32,768 tokens; played on 2026-10-26. Costs are what OpenRouter billed (list prices $1 per million input tokens, $5 output, $0.05 cached input, $1.25 cache write, half of Sol's). Its thinking parts are OpenAI's reasoning summaries, not the full reasoning; 46 of its 97 replies carry one.
• Benchmark: Large Barren Desert construction, weapons and armour in the armory counted at the rules' prices; version 1.7.3 (commit 55bcc41, 1.7.3 relocked with a retry for a reply cut off by the provider closing the connection ("terminated"); play is unchanged); no inference_retry event is in events.jsonl or logs.jsonl and no request failed. From the save the human reference game started from; 2,000 game seconds (33 min 20 s) at speed 40; no playbook; host on the game machine. Kimi K3.1 Pro's games played this commit, the other models' earlier 1.7.3 ones.
• Run 00c22e8c, series 20261026T202811-8cb67492, 2026-10-26: game 3 of GPT-6.1 Terra's six, a fresh series with no memory of any other, valid on its first attempt. Games 1 and 2 (Oasis by the Sea, with and without arms) were played earlier that day; game 4 (this map, no arms) and games 5 and 6 (The Sinkhole, with and without arms) followed it. All six games were played on 2026-10-26 (game 6 ended just after midnight UTC, on 2026-10-27). This game counts on the website's leaderboard.

Result

A full-budget score (all 2,000 game seconds). Net worth grew by 2,367 over leaving the game alone, 49.4% of the person's 4,792. By the website's formula (80% net-worth share, 10% population, 10% popularity; shares from doing nothing, 5,192 and 4 people, to the person, 9,984 and 250) the score is 50.8: 39.5 points from net worth, 6.0 from population (147 of the person's 246 people above the 4 at the start) and 5.3 from popularity. The stored goods are 245 stone (1,715), 310 bread (1,240), 46 iron (1,058), 24 crossbows (720), 17 flour and 3 wood; gold is 35% of net worth, stone 23%, bread 16%, iron 14%, crossbows 9.5%. The town ended on high taxes (150 gold a month), full rations, 24 idle peasants, no open job and 3 free housing places.

How the game went

Times are the recording's game clock when a call began; the run clock in the observations, the source of the idle, popularity and production readings quoted, reads a few seconds later. The game is paused while the model replies.

Opening, 0:00 to 3:00. map_overview showed no farmland or trees. The first build call (26 s) placed the marketplace, two hovels and four iron mines, every site find_sites offered; three stockpiles came at 0:31 and an armory, four fletcher's workshops and two hovels at 0:50. Its 0:31 summary: "Crossbows have a good buy and sell value, so they seem worthwhile." It built a chapel (2:11), set average taxes (2:25) and, with 24 people and 44 food, bought 75 apples, 75 cheese and 100 planks (2:33 to 2:57, 1,600 gold). The first payday came at 2:37 (9 gold), the first iron at 3:16.

Quarries, 3:42 to 10:00. Six more workshops made ten (3:58, 7:28). Eight quarries and 13 tethers went up between 4:41 and 9:05; the first stone arrived at 10:33. It bought 100 meat and 25 bread, sold iron, pitch and stone for 1,240 gold (8:12 to 8:25) and had 68 people at 9:00, 24 of them idle (first reached at 8:50).

The bread chain, 9:40 to 18:00. At 9:40 it first looked up the mill and bakery, bought wheat (10:03) and placed a mill and two bakeries (10:06); twelve bakeries stood by 17:23, still one mill. Its 13:44 summary: "10 wheat can produce 80 bread, which is quite profitable." Four chapels stood by 12:38; religion read +2 from 13:43.

Jobs by ox tether, 18:00 to 30:05. The rules say new peasants stop arriving while 24 stand idle; idle read 20 or more in 49 of 75 readings and open jobs were 0 in all but five. Its 24:58 summary: "I need more jobs since I’ve got 24 idle right now". From 18:32 it placed 34 more tethers in seven calls (63 in all) and its last hovel (20:32); taxes ran at high (18:02), mean (20:30), high (22:02). The one compaction came at 21:05.

The last 7 minutes, 26:47 to 33:20. Five waits (45 to 81 s), four tax changes and one build call (five tethers, 30:05). The bought foods ran out (cheese by 28:52, apples and meat by 29:52), the variety factor (+3 from 7:50) fell to 0 and popularity went from 66 to 53. Gold rose from 947 to 2,653.

The dip and the climb. On the recording's count (timeline below) net worth went from 4,661 (3:00) to 1,649 (18:00) and ended at 7,559. The game's own figure, summed from the observations' goods lists (gold plus every good at its sell price), was 5,955 at 1:25, 2,051 at 16:26 (the low) and 5,539 at 28:43: below the do-nothing 5,192 for about 25 minutes. Early on the town earned little (first payday 2:37, first stone 10:33) while it bought food and planks at two to four times their sell price and spent 1,600 gold on workshops and chapels; after 16:26 taxes, quarries and crossbows outran the market loss. The remainder line is what is left over: the logs have no consumption figure.

Each column is the last reading at or before that minute. Before the end the net worth row counts gold, wood, stone, iron and food (at 4); crossbows, flour, wheat, pitch and ale are left out, so it reads about 800 low at 30:00.

Weapons

It built the armory and ten fletcher's workshops (four at 0:50, three at 3:58, three at 7:28) and set each to crossbows within about a minute and a half (ten set_workshop calls, 79.7 game seconds; no bow appears in any goods list). The ten made 74 crossbows, 1.0 to 3.3 a minute together in the readouts: 50 sold in eight sales for 1,500 gold and 24 held, worth 720, 9.5% of net worth and 12.0 of the 50.8 points. Its 25:31 summary doubted the venture: "my investment in fletchers isn’t profitable as the costs outweigh the gains". It kept all ten.

The town had no trees and bought 1,035 of its planks, so the last line applies: the wing came out 488 behind (8.1 points), and each workshop also gave one job. Sol's five fletchers made 34 bows (510), 163 below even in Sol's analysis; Luna's wing (thirteen fletchers and a blacksmith) held 525.

Beside its other games and the other models

Against the person's growth the desert game reached 49.4%, the Oasis games 60.2% and 62.9%; all three ended with 24 idle peasants.

Terra is third of twelve by score and net worth, first by population (Argon has 134), second by gold, tied eighth by popularity, first by turns and sixth lowest by cost. It is 24.8 points behind Sol, 12.7 behind Argon (810 of net worth, 13.5 points, less 17 more people and a point of popularity) and 13.3 above Muse Wonder. Its desert map score, the mean of its two unrounded game scores (50.8 here, 66.5 in game 4), is 58.7: third of twelve, behind Sol (87.0) and Argon (70.3) and 15.6 above Opus 5.5 (43.1). With 62.4 on Oasis and 62.4 on The Sinkhole, its overall score, the mean of the three maps, is 61.2: third of twelve, behind GPT-6.1 Sol (69.7) and Gemini 4 Argon (65.6); on The Sinkhole alone it is first of twelve (Sol 49.8, Argon 49.7).

Sol's town was smaller (12 hovels, 104 people) and 1,610 richer. By Sol's analysis it paid in about 3,900 of taxes (Terra 6,328), mined 84 iron and 424 stone (Terra 96 and 560) and made 510 of bows (Terra 2,220). Terra took in about 5,400 more but paid out about 7,000 more: market loss 3,605, building gold 700, and about 2,700 more of wood used and food eaten for 47 more people. On the recording's count its low was 1,649 (18:00), Sol's 3,185 (24:00).

What held it back

• The market bill. Purchases lowered net worth by 7,750 against sell prices (on this map 1,000 of net worth is 16.7 points): food 4,645, planks 3,105; Sol's cost 4,145 (table above). The map has no farmland or trees, so part was unavoidable, but 5,280 gold went on apples, cheese and meat. By my balance (55 food at the start, 730 bought, about 640 bread baked, 310 left) the town ate about 1,100 units, about 700 of them apples, cheese and meat at 8 gold, while the wheat chain made bread for about 4 (the bread figure is my sum of the production windows, not a game figure).
• Popularity bought with food, not buildings. The rules name food variety, ale from an inn and religion. Terra never called building_info for the Inn and no summary mentions an inn or ale. Sol's three inns, about 300 gold by my subtraction from Sol's analysis, held +6 against its high tax; Terra's four chapels (1,000 gold) read +2 and the variety factor ended with the stock. It ended at 53 with −4 coming, Sol at 51 with 0.
• Four iron mines. 96 iron (2,208). find_sites found no further site at 3:42, 8:50, 21:18 or 22:24, while map_overview at 22:24 still listed ore patches of 35, 18, 16 and 6 tiles; the logs do not show what blocked them. Sol and Luna also placed four.
• Tethers made jobs, not stone. The quarries are 93 to 106 tiles from the keep. Stone averaged 30 a minute from 15:00 to 21:05 with 20 to 39 tethers and 24 after, with 39 to 63. The 43 tethers after 15:00 cost 215 wood (about 860 gold) but population rose by 56 while the new buildings added 52 jobs, so the tethers did serve as jobs.
• Cash and housing. Gold stayed between 244 and 1,311 from 13:00 to 27:00 (warnings at 15:39, 18:25, 24:37). No hovel after 20:32 and nothing built after 30:05: the end had 151 of 154 housing, 24 idle, no job open.

What worked

• A town that grew with no farms. 151 people, the most on this board, in 18 hovels; 70 paydays brought in 6,328 gold, the first at 2:37, six and a half minutes before Sol's.
• The bread chain. 115 bought wheat (2,645 gold) became about 640 bread, about 4 gold a unit against the market's 8; 310 bread stood in store at the end.
• Spending popularity. It stayed at 74 or more until 25:31 on taxes of levels 5 to 7 and ended at 53, three above the 50 below which the rules say peasants leave.
• Placement. All 123 buildings asked were placed in 24 build_structure calls.

Market

The 73 replies add up to 12,585 gold bought and 5,845 sold; the recording adds a bread click (40 gold) and 15 stone (105), giving:

The gold closes: 5,000 (at 0:32, the starting package's gold in) + 6,328 taxes + 5,950 sales − 12,625 purchases − 2,000 for workshops and chapels = 2,653.

I checked every trade against the gold, stone, iron, wood and food steps in frames.jsonl. No sale began from 200 or more in stock (the largest was 109 stone), so the known 1.7.3 undercount did not arise. Seventy replies agree once paydays (every 26.7 game seconds) are removed; three do not:

• 9:40 bread. The reply was not_traded, 0 units. At 9:43 gold rose by only 16 (a payday of about 56 less a 40-gold click) and food by 5: one lot of 5 bread was bought, a click that landed in a payday frame. Its 21:05 handoff still calls it a failed purchase.
• 19:46 stone. Reply: 45 units for 315. Recording: 70 + 8 × 35 = 350, 50 units; the first click sold 10.
• 23:19 stone. Reply partly_traded, 10 units for 70. Recording: a second +70 at 23:23, 20 units for 140.

What the food bought. The 8,485 gold of food kept the stock at 127 or more after 3:00, with no stock-out; it carried the variety factor (+2 from 3:42, +3 from 7:50), which let taxes run at levels 5 to 7 with popularity at 74 or more until 25:31; and it fed 151 people, the most on this board. Population came from housing and jobs while popularity stayed above 50, so food was the condition, not the cause. Buying it at twice its sell price lowered net worth by 4,645 at the moment of purchase, 2,640 of that on the three variety foods; the food was then eaten.

Cost and speed

97 turns, 238 tool calls (73 market_trade, 49 find_sites, 30 wait_and_observe, 24 build_structure, 16 set_tax, 15 building_info, 10 set_workshop, 7 inspect_building, 14 others), 4.48 million tokens and $0.456 billed (preparation $0.026, play $0.421, compaction $0.008); the interview added $0.006. All 97 gameplay requests read the cache (97.9% of input tokens); the one compaction (21:05, 12.7 s) cut the context estimate from 76,428 to 45,472. The run took 38.8 real minutes, 13.6 waiting for the model (median reply 8.2 s, longest 13.5 s); Sol's took 35.2 and 11.7, Luna's 48.3 and 21.8. Against Sol's $0.602 it used 32% more tokens (4.48 against 3.38 million) at half the list price and cost 24% less. Four images reached the model; all 30 waits ran without a screenshot, nine fast-forwarded (407.7 game seconds in 148.3 real seconds).

Of the 2,000 game seconds, market_trade took 544 (27%; median call 6 s, longest 24 s), build_structure 537 (24 calls, median 22 s, longest 35 s), the 30 waits 504, the 41 default 5 s passes after acting calls 209 (10.5%), set_workshop 80, inspect_building 51, set_tax 37, expand_storage 20 and set_rations 9; the read tools took none. build_structure placed 123 buildings of 123 asked: 4.4 game seconds per building placed here (Large Barren Desert), against 5.4 in Terra's Oasis weapons game (118 of 127 in 637 s) and 5.3 in its Oasis no-arms game (139 of 150 in 733 s). The measure depends on map, model and host, so it is not set beside other models.

The interview

It put the gap down to scale and focus: one mill and twelve bakeries, with resources and workers spent on quarries, oxen and workshops, against the person's bread business with bought wheat, religion and ale; keeping food variety was expensive on a map with no farms. Playing again it would build imported-wheat bread early and much larger, add housing with jobs, test church, inns and bought ale, try fear-factor buildings and add tethers only on observed deliveries. It asked for per-building output, goods in transit, quarry backlog and a live net-worth total.

These match the game: 7,559 against 9,984, gold 2,653 against 8,574, stored goods 4,906, one mill and twelve bakeries, quarries 93 to 106 tiles out, the variety food, popularity 53, 24 idle. Its 13:44 finding that the wheat chain paid and its 25:31 doubt about the workshops fit the closing lesson (test profitability, scale the best chain sooner). Not mentioned: the crossbows made, the dip to 2,051 and the variety lapse at 28:52. It would "test" inns and fear-factor buildings; in the game it never looked up either. The person's figures come from the interviewer's text and I did not check them.

For the harness

• market_trade replies out of step with the game. Three replies differ from the recording (9:40, 19:46, 23:19). At 9:40 the model was told nothing was bought and carried that into its handoff.
• No net-worth reading. At 28:52 the model's summary put net worth near 6,000; the observation's gold and goods sum to 5,748.
• The game saved correctly at the end ("gpt-6-1 terra 1-7-3 00c22e", recorded in episode.json).