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 with no other fallbacks (all 148 billed requests went to OpenAI), reasoning medium, output limit 32,768 tokens. 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; 101 of its 144 replies carry one.
• Benchmark: The Sinkhole construction (no arms): weapons and armour in store count nothing 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 fired and no request failed). 45 min of game time at speed 40; no playbook; host on the game machine. Kimi K3.1 Pro's games played this commit, Sol's, Argon's and Luna's here an earlier one (e0e0eb4).
• Run 85c60e4d, series 20261026T225945-9a2d33f2, 2026-10-26 (from 23:00 UTC; it ended just after midnight): game 6 of 6, the last of Terra's six, each a fresh series with no memory of any other, valid on its first attempt. Its weapons-counted pair is game 5 (run bdb7fe05, 1.7.1, commit 07f2b7a, that benchmark's round), so the pair differs in version as well as in the rule. This game counts on the website's leaderboard.

Result

A full-budget score (2,700.1 game seconds used). Net worth grew by 23,682 over leaving the game alone, 60.1% 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 59.1: 48.0 points from net worth, 5.2 from population (263 of the person's 502 people above the 4 at the start) and 5.8 from popularity (episode.json records arms_counted: false). Gold is 83.6% of net worth. The town ended on mean taxes (320 gold a month), full rations, 24 idle peasants, no open job and 7 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 5:55, wood, housing, first mines. The ore lies 52 to 88 tiles east of the keep, the nearest grove 39 tiles south. The first call (49 s) placed two woodcutters, two hovels, an orchard, a dairy farm and the marketplace; planks were bought at 0:54 and 3:31, iron mines went up at 1:14 and 3:53 (three each), the first iron came at 4:16. Average taxes at 2:09 took popularity from 100 to 75. Net worth stood under the do-nothing 2,195 from 1:03 to 11:20.

5:55 to 17:00, pikes and wood. No taxes at 5:55; the armory and two poleturner's workshops at 6:53, two more at 8:09; idle peasants first read 24 at 9:17; a brewery, an inn and four woodcutters at 9:42; average taxes at 11:02. Ten placements failed 2 to 16 wood short, seven of them iron mines (11:11 to 15:44).

17:00 to 36:00, bread, iron sales, religion. A mill and three bakeries at 17:55; planks at 17:34, 18:49 and 22:28, the last purchase of the game; the first pikes sold at 21:16, a church at 21:41; the first stone arrived at 23:19 (quarry placed 15:44). expand_storage was refused at 24:42; it sold 100 iron at 24:58, 30:41 and 34:16. A cathedral at 25:57 brought religion to +3; high taxes at 31:27 took popularity from 78 to 62 by 36:00.

36:00 to 45:00, woodcutters and the finish. From 36:08 it placed 32 woodcutters, 4 tethers and 6 bakeries and sold planks from 36:55. With about four minutes left it set extra rations (40:36) and mean taxes (40:45), then full rations at 43:27; popularity ran from 70 to 58.

Each column is the last reading at or before that minute; mines and woodcutters count when the placing call returned. Before the end net worth counts gold, wood, stone, iron and food (at 4) as the recording logs them, leaving out flour, wheat, hops and ale (860 at the end).

Taxes, rations and popularity. Taxes began at none; rations were full except 13:37 to 18:40 and 40:36 to 43:27 (extra, +4 to +6 with variety). Each spell runs to the next call:

The 87 paydays brought 11,882 gold (2,847 by 25:00, 6,311 by 35:00), summed from the recording's gold steps with the two frames that shared a market click corrected (22:43, 38:43). From 10:00 to 40:45 popularity stayed between 62 and 97.

The end settings. Mean taxes paid 297 to 320 a month. At the 0.796 a head the average spell paid, the ten paydays would have brought about 2,065 against 3,049: roughly 980 gold (2.0 points) more, while popularity fell from 69 to 58. At 43:27 it went back to full rations with food at 89 and falling about 43 a minute; that removed the +5 food factor, and popularity fell 12 in the last 90 s (1.2 points).

Without weapons

At 6:53 it looked up the Fletchers', Poleturner's and Armory entries; its reasoning summary reads "the workshop is allowed". It built the armory and eight poleturner's workshops (two at 6:53, two at 8:09, three at 23:20, one at 26:48). set_workshop by tile took for six (pikes); a seventh at 9:30 came back not_a_workshop and was not retried, so two shops made spears. The eight made 106 weapons in 229 workshop-minutes (66 pikes, 40 spears) and sold 105 for 1,570 gold (pikes at 18, spears at 10).

Planks cost 4 to buy until 22:28 and brought 1 when sold from 36:55; the town sold 530 and ended with 50. The wing came out between 1.0 points ahead and 0.5 behind, nearer the first: 76 of the 106 weapons came after the last purchase, and the last four shops were built with 83 to 170 wood in stock. Four other models built none on this benchmark; Terra's desert no-arms game built sixteen and ended 532 behind.

Beside its other games and the other models

All six ended with 24 idle peasants and cost $4.07 and 830 turns together; game 5 played 1.7.1, the others 1.7.3. The no-arms game scored 2.6 above its weapons-counted pair on Oasis, 15.7 above on the desert and 6.7 below on the Sinkhole.

Against game 5 (same map; 1.7.1 there, 1.7.3 here). Same iron, fewer taxes:

Game 5 set mean taxes at 14:36 with popularity at 100 (it fell to 73 by 25:48), offset by extra rations from 7:06 and religion +3 from 13:18, on a bread chain begun at 11:08; its paydays reached 5,362 by 25:00 against 2,847 here. This game had religion 0 until 22:27 and too little food for extra rations, and ran average and low taxes through those minutes. The rest of the gap is 1,540 less lost to purchases and 500 less in buildings.

Terra is first of twelve by score (9.9 above Sol, 12.0 above Argon), net worth, population (86 above Argon) and gold, 11th by popularity (Luna's 57 is lower), first by turns, sixth from the lowest by cost.

Where the margin came from. Net worth gained since the start (2,415), by source, from each run's gold steps, goods and replies (Sol's and Argon's checked against their recordings):

Terra's remainder: weapons sold 1,570, food and flour made net of eaten 1,420, wood 25. Against Sol the 5,764 gap in net worth (11.7 of the 9.9 points) is taxes +3,949 and iron +3,427, less 1,680 stone and 2,150 more spent on buildings, with 500 less lost to purchases and 1,718 more remainder; population adds 2.4 points and popularity takes back 4.2. Against Argon the 5,773 (11.7) is taxes +3,068, iron +2,668 and remainder +2,017, less 2,016 stone (36 net from the rest); population adds 1.7 and popularity takes back 1.4. Iron per mine was alike (17.6, 17.1 and 16.4), so the gain is the number and timing of mines, which rested on wood: 20 woodcutters at 15 minutes against 3 and 5. The lead over Argon grew from 641 at 35:00 to 3,975 at 40:00 (Sol: 3,294 and 4,703).

Against Sol it used 57% more tokens (7.21 million against 4.60) at half the list price and cost 11% less ($0.731 against $0.825).

Its board position, complete. Map scores are the mean of a model's two unrounded game scores; overall is the mean of the three maps. Terra's Sinkhole score is 62.4 (65.8 and 59.1): first of twelve, 12.6 above Sol (49.8) and 12.7 above Argon (49.7). Its Oasis score (62.4) and desert score (58.7) are both third; overall it is third of twelve at 61.2, 8.6 behind Sol and 4.4 behind Argon, 22.7 ahead of Opus.

What held it back

• Popularity at the end. 58, 11th of 12, against Sol's 100 (4.2 points); mean taxes and full rations took it from 70 to 58 in the last 90 seconds.
• Wood, early. Ten placements failed for wood (seven iron mines, 11:11 to 15:44) while 13 to 20 woodcutters gave 36 to 63 a minute. The 500 planks cost 2,000 gold and 1,500 of net worth (3.0 points).
• Stone. Three quarries and 22 tethers (110 wood, 22 jobs) delivered 144 stone against Sol's 384 and Argon's 432.
• A thin bread chain. First bakeries at 17:55; 17 at the end (Sol 24, Argon 31, the person 161), 73 flour left unbaked; extra rations for 8 of 45 minutes.
• Religion and the inn for little. The chapel read religion 0 until the church (22:27); religion cost 1,750 gold (3.6 points of net-worth share) for +3 from 26:39. The inn read +2 in four readings (17:10 to 18:28) and 0 in the other 77; the brewery gave 0.3 to 0.7 ale a minute with 13 hops unused. It never inspected either.
• The idle cap. Idle read 24 in 22 of the 81 readings after 10:00, 20 or more in 68; no reading showed an open job.

What worked

• Iron. 29 mines mined 509 iron (11,707), 45.2% of the final net worth, at 0.61 a mine-minute from each placing call's return; 8 mines stood by 10 minutes and 17 by 20.
• Taxes. 11,882 gold over 87 paydays, 5,571 in the last ten minutes.
• Jobs. 267 people, 86 more than any other model, 243 at work. After 36:08 it placed 42 jobs (32 woodcutters, 4 tethers, 6 bakeries) and the number at work rose by 41; its 43:04 summary calls woodcutters "the fastest option". The 79 woodcutters cost 237 wood and cut most of the 1,845 wood the buildings used.
• Placement and handoffs. 239 of 253 buildings placed in 39 calls, each of the 14 refusals with a remedy. Three compactions cut the context estimate from about 78,000 to 45,000, and their figures matched the game.

Market

27 calls, 249 game seconds (9.2%), checked against the gold, wood, stone, iron and food steps in frames.jsonl, whose figures are used.

One reply is out of step: the planks at 39:52 began from 218 in stock; the recording shows five clicks (50, then four of 20), 130 sold for 130 where the reply said 100 and 100, its stock (88) and gold right. That is the 1.7.3 undercount for sales from 200 or more in stock. The four iron sales began from 175, 172, 130 and 100 and match their replies, as do the other 22 calls; all five plank purchases landed in full. Selling at the sell price leaves net worth unchanged for iron, stone, apples and planks; only the weapons add (1,570). Iron fell from 175 (25:00) to 24 (40:00) because 400 were sold against about 250 mined; nothing consumed it. The gold closes: 2,000 + 11,882 taxes + 12,400 sales − 2,000 purchases − 2,650 for buildings (workshops 800, inn 100, religion 1,750) = 21,632. No gold was spent after 27:34. It sold to free storage space, which never ran out (2 empty piles at the fewest).

Cost and speed

144 turns, 295 tool calls (164 find_sites, 39 build_structure, 27 market_trade, 65 others), 7.21 million tokens and $0.731 billed (preparation $0.026, play $0.678, compaction $0.027); the interview added $0.006. All 144 gameplay requests read the cache (97.5% of input tokens). The run took 59.4 real minutes, 18.5 waiting for the model. 101 replies held one call and 43 held three to seven; find_sites came in 45 turns. Seven images reached the model in play.

Of the 2,700 game seconds, build_structure took 1,624 (60.2%; median call 44 s, longest 76 s), the 98 default passes after acting calls 492, market_trade 249, inspect_building 148, set_workshop 53 and the two waits 68 (one fast-forwarded: 30 game seconds in 11 real). It placed 239 of 253 buildings: 6.8 game seconds per building placed here (The Sinkhole), against 7.2 in game 5 (The Sinkhole, 1.7.1), 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.

The interview

It put the gap down to the economy's design: taxes and generous rations where the person used inns, religion and fear, 79 woodcutters and 17 bakeries against 7 and 161, and "a low-value raw material". Playing again it would build the bread chain, ale, inns and religion earlier and add housing and cheap jobs together.

These match the game: the totals, 79 and 17 against 7 and 161, and taxes as a main source (11,882 of the 24,282 gold taken in). These do not. It says it "treated it too conservatively" (popularity), but it spent it and ended at 58, 11th of 12. It would "test imported ale and inns early", yet it had built an inn (paid 10:55), a brewery, four hops farms, a chapel (religion 0 for 17 minutes), a church and a cathedral, and bought no ale. It would "test fear-factor buildings"; it never looked one up. It reads the rules as prohibiting "unproductive military construction"; they bar soldiers and defences, and the workshops it built are allowed. The delivered output by building and the worker efficiency it asks for are in every observation (efficiency 100% in all); consumption is not.

For the harness

• market_trade undercount (39:52): the plank sale from 218 sold 130 where the reply said 100; the model did not notice.
• set_workshop miss (9:30): not_a_workshop asked for a pixel from a screenshot; the model had taken none and left that shop on spears.
• Far ore: two mines came back not_on_screen (344,581 at 28:45; 288,278 at 34:44), each beside one placed in the same call; find_sites had listed both tiles, and 288,278 also failed for Kimi K3.1 Pro and GPT-6 Luna.
• expand_storage (24:42, 32:20) said no free ground touches the stockpile or granary; storage never filled.
• No provider errors, stalled replies or refused calls, and no forced collection by the reader watchdog (forced 0 in results.json). The game saved at the end ("gpt-6-1 terra 1-7-3 85c60e").