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 156 billed requests, 1 preparation, 153 gameplay and 2 compaction, were billed to OpenAI), reasoning medium, output limit 32,768 tokens; played on 2026-10-26. Costs are what OpenRouter billed (list prices $1 and $5 per million input and output tokens, half of Sol's). Its thinking parts are OpenAI's reasoning summaries, not the full reasoning; 74 of its 153 replies carry one.
• Benchmark: Large Barren Desert construction (no arms): weapons and armour count nothing in the armory and bring their gold if sold; 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 unchanged; no inference_retry event is in events.jsonl or logs.jsonl and no request failed, so it did not fire). 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' 1.7.3 games earlier ones.
• Run fd082c2a, series 20261026T210849-99fcf8d0, 2026-10-26: game 4 of GPT-6.1 Terra's six, a fresh series with no memory of any other, valid on its first attempt. Games 1 to 3 (both Oasis games, then this map with weapons counted) were played earlier that day; 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 (2,000.1 game seconds used). Net worth grew by 3,247 over leaving the game alone, 67.8% 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 66.5: 54.2 points from net worth, 6.4 from population (158 of the person's 246 people above the 4 at the start) and 5.9 from popularity. episode.json records arms_counted: false. Gold is 59.7% of net worth, stone (262, 1,834) 21.7%, iron (53, 1,219) 14.4%; no spear is stored. The town ended on mean taxes (194 gold a month), full rations (4 of 4 foods eaten, 18 food stored), 24 idle peasants, no open job and no free housing.

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.

Opening, 0:00 to 3:24. map_overview showed no farmland, oil or trees, stone 83 to 106 tiles west of the keep and iron 57 to 69 tiles east. The first build call (26 s) placed the marketplace, two hovels and the four iron mines find_sites offered; three stockpiles, a chapel, three quarries and four more hovels followed by 1:25. It set high taxes (2:02), bought 30 each of cheese, apples and bread (2:10 to 2:34, 720 gold) and 100 planks (3:02). The first payday came at 2:10 (4 gold), the first iron at 3:17.

The spear works, 3:24 to 5:28. An armory and seven poleturner's workshops at 3:24, five more at 4:15 and four more with an inn and two hovels at 5:28: sixteen, 1,600 gold. Gold fell from 4,055 (3:02) to 1,840 (6:18); the first sale was 15 spears at 7:35.

Stone and church, 6:17 to 15:00. Mean taxes came at 6:30, the first stone at 7:27, the church (500 gold) at 8:04. Idle peasants read 20 at 8:55 and 24 from 15:10, the number at which the rules say new peasants stop arriving; its 12:59 summary had said "my growth is capped at 24" and eight tethers followed.

Inns, tethers and the bread test, 14:30 to 29:37. Inns three and four came at 14:30 and extortionate taxes at 15:51. A mill and four bakeries came at 16:46; it bought 10 wheat at 17:10 and 10 at 20:51, the only wheat of the game. Nine tether calls from 15:23 to 29:37 added 44 (68 in all), with quarries at 19:27 and 22:07 (11 in all); the last hovel came at 27:01.

The last 4 minutes, 29:20 to 33:20. Taxes went 7, 6, 7 (29:21, 30:49, 31:06) and rations half, then full (29:29 to 30:58). Nothing was built after 29:37 (162 of 162 housed, 24 idle, no open job in any reading after 4:40 but one). The last purchases were small (31:25 to 31:55); a food-low warning came at 32:51 (46 left, 29 s to go); the last spears were sold at 33:09.

Each column is the last reading at or before that minute; before the end the net worth row leaves out ale, pitch, flour and wheat (about 300 at 30:00).

The dip and the climb. By the game's own figure, summed from the observations' goods lists (gold plus every good at its sell price, spears at 0), net worth was 5,824 at 2:33, under the do-nothing 5,192 from 3:59 to 27:36 and lowest at 2,381 (15:23); from there it rose 6,058. The logs have no consumption figure, so the remainder is what is left over.

Without weapons

The rules say weapons "count nothing" in the armory and "sold at a marketplace, it brings its gold". It built an armory and sixteen poleturner's workshops to sell spears. They make spears unless told otherwise, so no set_workshop call was needed. Its 1:25 reasoning summary planned "profitable workshops for converting wood into bows and other weapons"; none says why poleturners rather than fletchers (game 3 used fletchers).

The sixteen made 288 spears, all sold in 13 sales (2,880 gold, each sale agreeing with the recording's gold steps; never more than 30 in the armory). The readout gave 8.7 to 11 a minute together from 11:35, about 0.6 a workshop; wood stood at 2 or less for only 24 s after 5:00, so the workshops, not wood, set the pace.

The town bought 1,115 planks and ended with 3, so the right-hand column applies: the wing came out 532 behind (8.9 points). At 4 a spear brought 6 and the 2,260 of fixed cost needed 377 spears; at 1 the wing paid back between 23:52 and 26:28. By 15:23 it accounted for −1,009 to −1,756 of the 3,443 fall in net worth since 2:33. Each workshop was also one job, 16 of the 138 working people; a tether is a job for 5 wood and the model placed 68, so 16 tethers cost 80 wood. Its 19:27 summary put the payback at 11 minutes, "but that feels marginal given taxes"; it looked up workshop sites again and built none. Sol built no weapon works here; Luna sold 14 maces (420 gold).

Beside its other games and the other models

Games 5 and 6 (The Sinkhole, with and without arms) were played after this one. All six games ended with 24 idle peasants. Over the three maps its scores are Oasis 62.4, desert 58.7 and The Sinkhole 62.4, and 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).

Against game 3 (same map, model and day) the score is 15.7 higher: net worth +880 (+14.7 points), 11 more people (+0.4), 6 more popularity (+0.6). The 880 is gold +2,388 and goods −1,508 (bread −1,192, crossbows −720, flour −170, ale +270, stone and iron +280, other food +24); the gold is taxes +3,233 (9,561 against 6,328), sales +660, purchases −755 and buildings −750. Game 3 had four chapels, no inn and 12 bakeries on 115 bought wheat; this game has a chapel, a church, four inns and 4 bakeries on 20 wheat, and ended on mean taxes (194 a month) against game 3's high (150). Game 3's end would score 38.8 under this rule (6,839 without its 720 of crossbows); this end holds no weapon, so it scores the same either way.

Terra is fifth of twelve by score and net worth, first by population (Argon has 136), second by gold, first by turns (Sol 77) and sixth from the lowest by cost. It is 0.6 behind Opus (369 less net worth, 6.2 points, against 5.6 for 63 more people and 30 more popularity), 10.5 behind Argon and 32.0 behind Sol. Its desert map score, the mean of its two unrounded game scores, is 58.7: third of twelve, behind Sol (87.0) and Argon (70.3), ahead of Opus (43.1) and DeepSeek (39.6).

Sol's town was 57 people smaller and 1,898 richer: it sold nothing and ended on 570 stone and 104 iron (6,382) at low taxes and 85 popularity. Terra mined as much (552 stone and 88 iron against 520 and 89) and took 6,549 more in taxes, but lost 5,080 more net worth in purchases and spent 2,550 more on buildings.

What held it back

• The market bill. Purchases lowered net worth by 7,875 against sell prices at the moment of purchase (on this map 1,000 of net worth is 16.7 points): food 3,780 (945 units at 8 gold, then eaten), planks 3,345, ale and wheat 750.
• The spear works. 532 gold and 8.9 points behind, a third of the wood it used (453 of 1,347), and 1,600 gold out between 3:24 and 5:28.
• The bread chain stayed at 20 wheat. The four bakeries delivered about 160 bread (my sum of the readout's windows; the food balance after 18:09 agrees). That is about 8 bread a wheat, twice the 4 of its 30:48 handoff: 460 gold of wheat replaced about 1,280 of bought bread. It bought no wheat after 20:51 and the bakeries delivered nothing in the last three minutes.
• Four iron mines. 88 iron (2,024) was mined. find_sites found nothing within 25 tiles at 0:49, 15:23 and 19:27 while map_overview lists four ore patches (62, 25, 18 and 16 tiles), two holding the mines; the logs do not show what blocked the rest. Sol and Luna also placed four.

Would the town have run short? Yes, within about 20 seconds of the end: food read 189 (27:00), 153 (30:00) and 18 at the end, falling 56 to 59 a minute over the last 60 s with the bakeries idle. It ran the stock down on purpose; its 31:45 summary counted seconds of cheese to the end.

What worked

• A town that grew with no farms. 162 people, the most on this board; 71 paydays brought 9,561 gold. With 24 idle from 15:10, each added person came with a job: 63 jobs from 44 tethers, 4 quarries, the mill and 4 bakeries for 63 more people (99 to 162).
• Spending popularity. Four inns (400 gold, 80 wood, 45 ale for 900) carried the inn factor to +8 and taxes ran at level 8 from 15:51 to 29:21: a payday of 222 at 159 people against 190 to 194 at level 7. Its 26:28 summary worked out a fall of 28 from 79 to 51 by the end; popularity read 59 with no change from 31:07 to the end, above the 50 below which peasants leave.
• Placement and cash. 131 of 133 buildings were placed in 24 calls, both refusals (10:07, 10:41) naming the remedy. Gold never fell below 892 after 3:00.

Market

The 74 replies add up to 13,340 gold bought and 6,610 sold; the recording adds one lot of 5 apples (40 gold) that the 27:53 reply missed, giving:

The gold closes: 5,000 (the starting package, in by 0:32) + 9,561 taxes + 6,610 sales − 13,380 purchases − 2,750 for buildings (workshops 1,600, church 500, inns 400, chapel 250) = 5,041. The 9,561 is the recording's 71 paydays (4 to 222 gold): 9,491 as read, 9,561 with the clicks that shared a payday's frame put back.

I checked every trade against the gold, wood, stone, iron and food steps in frames.jsonl. No sale began from 200 or more in stock (the largest was 118 stone), so the 1.7.3 undercount did not arise. One reply is out of step, against three in game 3:

• 27:53 apples. The reply was not_traded (0 clicks, "did not show in the reader within 5 seconds") after a 24 s call, the longest of the 74. The recording shows a payday of +217 and, 0.3 s later, −40 with food +5: one lot of 5 bought. Apples fell 4 by 28:31 while cheese and meat fell 9 each, which fits it. The model observed (28:22) and bought 25 more.

In five other calls a click shared a payday's frame (9:35, 12:47, 25:38, 26:04, 31:25) and each reply still matches; the other 73 agree.

Cost and speed

153 turns, 173 tool calls (74 market_trade, 36 find_sites, 24 build_structure, 39 others), 7.43 million tokens and $0.640 billed (preparation $0.026, play $0.596, compaction $0.018). All 153 gameplay requests read the cache (98.4% of input tokens). Two compactions (16:45 and 30:48, 27 s together) cut the context estimate from about 75,700 to 45,500. The run took 46.1 real minutes, 18.7 waiting for the model (median reply 7.0 s, longest 19.3 s); Sol's took 37.6 and 13.1, game 3's 38.8 and 13.6. Against Sol it used twice the tokens at half the list price and cost 7% less; against game 3, 66% more tokens and 40% more cost. 148 of its 153 replies carried one call (game 3: 57 of 97); five carried 4 to 6 read-only calls. Its 6:17 summary says the trades use the same interface and "it has to be one at a time".

Of the 2,000 game seconds, market_trade took 621 (31%; median call 7 s), build_structure 556 (24 calls, median 22 s), the 117 default 5 s passes after acting calls 591 (29.6%; game 3's 41 took 209), the 4 waits 105 and the other tools 120. build_structure placed 131 of 133 buildings asked: 4.2 game seconds per building placed here (Large Barren Desert), against 4.4 in game 3 (Desert) and 5.4 and 5.3 in games 1 and 2 (Oasis by the Sea); the measure depends on map, model and host.

The interview

It put the gap down to the bread chain: it "did not scale" it and tested it "too late" and "at too small a scale". Playing again it would build mills and bakeries in stages from the start, add religion and inns, test fear-factor buildings, measure quarry deliveries before building tethers and sell all weapons.

These match the game: 8,439 against 9,984 (15.5% less), 162 against 250 people, 59 against 100, a small and late wheat test, 945 food units bought, every spear sold. It says its economy "leaned on quarries, iron and spear sales"; taxes were the larger source, 9,561 of the 16,171 gold taken in, spear sales 2,880. It does not mention the loss on the workshops or the dip to 2,381, and it would "test" fear-factor buildings that it never looked up. The person's figures come from the interviewer's text and I did not check them.

For the harness

• One market_trade reply out of step. 27:53 told the model nothing was bought when 5 apples were; its 30:48 handoff records "observation confirmed no stock increase", which the stock figures above do not bear out.
• Reader watchdog. One forced collection at 22:07 after a 20.1 s stall (forcedCollections 1 in episode.json); the game stayed up.
• Warnings (1.7.0) fired once, food low at 32:51. The game saved correctly ("gpt-6-1 terra 1-7-3 fd082c").