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 145 billed requests, 1 preparation, 141 gameplay and 3 compaction, went 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, $1.25 cache write, half of Sol's; the gameplay bill matches them). Its thinking parts are OpenAI's reasoning summaries, not the full reasoning; 79 of its 141 replies carry one.
• Benchmark: The Sinkhole construction, weapons and armour in the armory count at the rules' prices; version 1.7.1 (commit 07f2b7a), that benchmark's round: the runner checked out 07f2b7a for it, like the other models' listed games there, so Terra's other games (1.7.3, commit 55bcc41) ran a newer harness. 1.7.1 has no retry for a reply cut off by the provider closing the connection ("terminated"); none was needed (no inference_retry, no failed request). From the save "The Sinkhole start"; 45 min of game time at speed 40; no playbook. Most models' 1.7.1 Sinkhole games were hosted on the Mac; Fable's, DeepSeek's, Kimi's and Terra's on the game machine.
• Run bdb7fe05, series 20261026T215649-b35cbfef, 2026-10-26: game 5 of GPT-6.1 Terra's six that day, each a fresh series with no memory of any other, valid on its first attempt. Games 1 to 4 (both Oasis games, both desert games) were played earlier that day; game 6 (The Sinkhole, no 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 (2,700.2 game seconds used). Net worth grew by 25,862 over leaving the game alone, 65.6% of the person's 39,436. By the website's formula (80% net-worth share, 10% population, 10% popularity; shares from doing nothing, 2,195 and 4 people, to the person, 41,631 and 506) the score is 65.8: 52.5 points from net worth, 5.6 from population (281 of the person's 502 people above the 4 at the start) and 7.7 from popularity. It is the best of the twelve weapons-counted Sinkhole games on the board, 13.4 above Gemini 4 Argon (52.4) and 15.3 above GPT-6.1 Sol (50.4). Gold is 62.3% of net worth, iron (287, 6,601) 23.5%, bread (489, 1,956) 7.0%, the 34 stored weapons (765) 2.7%. The town ended on no taxes and double rations, with no open job and 5 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.

0:00 to 14:36. The first call (41 s) placed three woodcutters, an orchard, a dairy farm, two hovels and the marketplace; 100 planks were bought at 0:46 and 2:10, three iron mines went up at 1:07. No farmland lay in view, so orchards went 23 to 146 tiles from the keep (median 75) and the 29 iron mines 41 to 179. Average taxes came at 3:39, the armory and first workshops at 3:48 and 4:47, a chapel at 6:06. Popularity fell to 79 by 7:03; extra rations (+4) followed at 7:06 and a church at 9:38 brought religion to +3 from 13:18. Idle peasants first read 24 at 7:39; the bread chain began at 11:08. Net worth stood under the do-nothing 2,195 from 1:07 to 14:25 (low 1,561 at 10:45).

14:36 to 35:34. Mean taxes at 14:36 with popularity at 100; 22 mines by 17:45; first bread 17:45; quarries 19:05 and 20:10 (first stone 25:28). Popularity fell to 73, so high taxes at 25:48. A third church, four mines 126 to 136 tiles south (28:47), a second granary and far-east dairies followed. Net worth passed Argon's and Sol's between 20 and 25 minutes.

35:34 to 45:00. Mean taxes again (popularity 80 to 70 by 38:27), then high; bakeries went from 20 to 39 (35:42 to 40:24). From 41:29 it put up 24 woodcutters in five calls; the 42:06 handoff says "new woodcutters are cheap jobs and can keep peasants arriving if idle count approaches 24". Idle read 15 at 41:16 and 23 at 44:47; population went from 254 to 285.

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) as the recording logs them; flour, wheat and weapons are left out (about 1,000 at 40:00).

Taxes, rations and popularity. Rations were full to 7:06, extra (+4) to 42:59, double (+8) to the end; half rations were never used. Taxes began at none; each spell runs to the next call:

The 93 paydays brought 15,967 gold (5,362 by 25:00, 9,800 by 35:00), summed from the recording's gold steps with the clicks that shared a payday's frame put back. It raised the rate while popularity stood near 100 and cut it at 73 and 70; the 16:17 handoff says to "reduce tax if popularity approaches 70". The first payday after a change is a blend (43:10 281, then 382).

The end settings. Double rations (42:59) cost no stock that showed: bread rose from 402 to 490 while apples ran out. Extortionate taxes (43:07) paid 382, 388 and 394 against about 270 at high taxes, some 370 gold more, while popularity slid 76 to 69. At 44:36 the reasoning summary reads "I have about 29 seconds to switch to no taxes, and I might gain around 11"; popularity rose 69, 72, 75, 77 in the 21 seconds left and the last payday (44:57) brought 99 against the 394 before it. Extortionate taxes took popularity from 76 to 69 for about 370 more gold; the switch to none gave back 8 of popularity (0.8 points) for about 300 less gold in the last payday (0.6 points).

Weapons

It built an armory (3:48) and 14 fletcher's workshops (4:47 to 32:14): 1,400 gold, 280 wood, 5 wood for the armory. Four were set to crossbows by tile (7:25, 7:38, 7:53, 33:17; all took), ten made bows. At the rules' prices (bows 15, crossbows 30) they made 67 bows and 37 crossbows, 2,115, from 245 wood. 70 were sold for 1,350 gold (20 bows at 28:14, 20 crossbows at 32:02, 30 bows at 39:22) and 34 are held (17 and 17, 765): 2.7% of net worth, 1.55 of the 65.8 points (64.2 with weapons worth nothing).

At wood's sell price of 1 the wing returned 2,115 for 1,400 gold and 530 wood, +185 (0.4 points); at the 4 gold it paid for planks, −1,405 (−2.9). At 43:54 the readout gave 4 weapons a minute for 14 workshops (about 6 gold each) against 15 iron for 29 mines (about 12 each). The 3:39 summary planned "converting wood into bows"; at 33:51 it wrote "To break even, I'd need around 28 minutes if the output is only 0.2 per building" and built no more. The wing covered itself at best; each workshop was also one job at the idle cap.

Beside its other games and the other models

All six ended with 24 idle peasants.

Terra is first by score, net worth, population and gold, third by popularity, second by turns and sixth lowest by cost. Its Sinkhole map score, the mean of its two unrounded game scores (65.8 here and 59.1 in game 6, no arms, on 1.7.3; game 6 is also first of the twelve no-arms games, Sol's 49.2 next), is 62.4: first of twelve, ahead of Sol (49.8) and Argon (49.7). Sol's map scores on the other two maps (72.3, 87.0) are above Terra's (62.4, 58.7); Terra's 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).

Where the margin came from. The net worth gained since the start (2,415), by source, from each run's own gold steps, goods and replies:

Against Argon the 4,868 gap in net worth (9.9 of the 13.4 points) is taxes +4,318 and iron +3,749, less 784 stone, 1,080 more lost to purchases and 2,450 more gold in buildings, with 1,115 remainder; population adds 1.4 points and popularity 2.1. Against Sol the 7,673 gap (15.6 points) is taxes +6,053 and iron +8,349, less 1,050 stone, 1,940 purchases, 2,650 buildings and 1,089 remainder; population adds 2.1, popularity −2.3. Terra taxed from 3:39 while they ran low taxes to 20 minutes, collecting 5,362 gold by 25:00 against about 2,250 and 2,150, with 174 people against 127 and 115. It paid for taxes in popularity (down to 73) where Sol held popularity at 100 on double rations and low taxes. GPT-6 Luna (26.7) had 24 mines, 9 woodcutters, no bakery and 105 people.

What held it back

• Woodcutters. 69 of the 256 placements, 28 after 40:00, yet wood income stayed flat: 13 woodcutters made 39 a minute at 9:39, 45 made 46 at 41:16. It bought 880 planks for 3,520 gold (net worth −2,640, 5.4 points), and 7 of the 13 unplaced buildings were 1 to 20 wood short. The 28:26 handoff says "no diagnosis established".
• The idle cap. Idle read 24 in 23 of the 75 observations after 10:00 and 20 or more in 55 of the 68 from 7:39 to 41:00; open jobs read 0 but once, with 4 to 40 housing places free. The 16:17 handoff says "24 idle blocks further immigration".
• Popularity spent, not built. It ran 70 to 80 under mean and high taxes with no inn or ale; the chapel and three churches (1,750 gold) gave religion +3 in 15 readings (13:18 to 23:41) and +1 in the other 53 after 10:00.
• Bread late. First bakeries at 11:08; 19 of the 39 (with 11 wheat farms and 4 mills) came after 35:42; 489 bread stood unsold.
• Storage and build time. expand_storage was refused for the armory (25:27) and stockpile (35:02); build_structure took 68.6% of the game clock.

What worked

• Population early. 47 people at 5:00 against 43 and 33; 141 at 20:00 against 104 and 98. 40 orchards and dairies fed them without a mill; food never read under 28.
• Taxes under a floor. 15,967 gold; popularity never fell under 69 after 10:00 (Argon's low was 54).
• Iron. 29 mines delivered 502 iron (11,546), 41% of the final net worth, at 0.5 to 0.6 a mine a minute.
• Placement. 256 of 269 buildings placed in 38 calls; each refusal came with a status and a remedy.
• Handoffs. Three compactions (16:17, 28:26, 42:06) cut the context estimate from about 76,000 to 47,000.

Market

19 calls, 238 game seconds (8.8% of the budget). Every trade was checked against the gold, wood, stone, iron and food steps in frames.jsonl, whose figures are used.

One reply is out of step: the iron sale at 35:21. It asked for 20 lots from 314 in stock; the reply said 100 units and 2,300 gold; the recording shows five clicks, three of 50 (1,150 each) and two of 20 (460): 190 iron and 4,370 gold (gold 6,027 to 10,397, iron 314 to 131). That is the 1.7.1 undercount for sales from 200 or more in stock; the reply's stock and gold were right. The model saw it next turn ("I sold 190 units instead of the expected 100") and wondered about duplicate sales. It is the gold rise from 5,805 (35:00) to 14,173 (40:00) while iron fell from 312 to 204: +8,368 is the iron sale 4,370, stone 700, bows 450 and taxes 2,848. A sale leaves net worth unchanged, so the score was not affected. Every plank and cheese click landed; the other 18 replies match.

The gold closes: 2,000 + 15,967 taxes + 6,995 sales − 4,320 purchases − 3,150 for buildings (14 fletcher's 1,400, three churches 1,500, chapel 250) = 17,492.

Cost and speed

141 turns, 314 tool calls (194 find_sites, 38 build_structure, 24 building_info, 19 market_trade, 38 others), 7.05 million tokens and $0.765 billed (preparation $0.026, play $0.706, compaction $0.032); the interview added $0.007. All 141 gameplay requests read the cache (97.4% of input tokens). The run took 61.0 real minutes, 19.0 waiting for the model (median reply 7.7 s, longest 15.0 s) and 22.9 inside build_structure (Sol's run took 53.5, Argon's 61.2). Against Sol it used 36% more tokens at half the price and cost 15% less.

find_sites came in 50 of the 141 turns, up to 8 at a time (94 turns held one call, 48 only read calls): 0.76 searches per building placed, no game time, 6.2 real minutes. Nine images reached the model all game: the first screenshot, three map overviews, a flat view (two) and three screenshots after compactions.

Of the 2,700 game seconds, build_structure took 1,853 (68.6%; median call 51 s, longest 74 s), the 83 default passes after acting calls 418 (15.5%), market_trade 238 and the other tools about 200 (the measures overlap by about 15 s). It placed 256 of 269 buildings: 7.2 game seconds per building placed here (The Sinkhole), against 5.4 and 5.3 in games 1 and 2 (Oasis by the Sea) and 4.4 and 4.2 in games 3 and 4 (Large Barren Desert); the measure depends on map, model and host. Calls of 7 or 8 placements took 33 to 74 s, a single placement 9 to 10 s.

The interview

It put the gap down to economic design: taxes and generous rations where the person used inns and religion, a much larger bread chain and housing, and "68-plus" woodcutters against seven. Playing again it would build the bread chain earlier, test ale, inns and religion, add housing and jobs together, use fewer woodcutters and sell surplus regularly.

These match the game: 28,057 is 67% of 41,631 and 285 is 56% of 506; 69 woodcutters against 7; extra rations for 36 minutes and taxes the larger share of the gold taken in (15,967 of 22,962); the bread chain from 11:08; two refused expansions; building taking most of the clock. Two points need a note. It would "test" inns and religion early: it built a chapel (6:06) and three churches (9:38, 17:45, 25:56), but never looked up an inn or a brewery. It says the screenshots helped, though 9 images reached it in 141 turns. Fear-factor productivity and water pots come from the interviewer's text; it never searched for either.

For the harness

• market_trade undercount (35:21). The reply showed about half the iron sold; the model found it from the gold. The replies' sold totals add to 4,925 gold, the recording's to 6,995.
• No room beside the armory or stockpile (25:27, 35:02), as in Sol's, Sonnet's, GLM's and Kimi's 1.7.1 games. The reply says to sell or demolish; the model sold weapons, iron and stone five times.
• Far sites. Four not_on_screen refusals (orchard 19:05, dairies 29:47 and 32:14, mine 33:51) carried the remedy to pick an inward tile; the two dairies were in the same far-east corner, 2½ minutes apart.
• One forced collection by the reader watchdog before timed play; set_workshop by tile worked on all four calls. The game saved at the end ("gpt-6-1 terra 1-7-1 bdb7fe").