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 160 billed requests were served by OpenAI), reasoning medium, output limit 32,768 tokens; released on OpenRouter and 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). Its thinking parts are reasoning summaries, not the full reasoning.
• Benchmark: Oasis by the Sea construction, weapons and armour in the armory counted toward net worth, version 1.7.3 (commit 55bcc41, 1.7.3 relocked: it adds a retry for a reply cut off by the provider closing the connection ("terminated"), with the game paused; what the model plays is unchanged), from the save the human reference game started from; 33 min 20 s of game time (60,000 ticks) at speed 40; no playbook; host on the game machine. One inference_retry fired, for a different error (see Cost and speed). Kimi K3.1 Pro's games played this commit.
• Run 8b331985, series 20261026T184310-000b1db6, 2026-10-26: game 1 of GPT-6.1 Terra's six, a fresh series with no memory of any other, valid on its first attempt. Game 2 (Oasis, no arms) and games 3 to 6 (the desert and the Sinkhole, each 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: the run used all 2,000 game seconds. Net worth grew by 11,010 over leaving the game alone, 60.2% of the person's 18,293. By the website's formula (80% net-worth share, 10% population, 10% popularity; shares from doing nothing, 1,220 and 4 people, to the person, 19,513 and 202) the score is 61.1: 48.1 points from net worth, 6.5 from population (128 of the person's 198 people above the 4 at the start) and 6.5 from popularity. The stored goods are 271 iron (6,233), 30 crossbows (900), 117 stone (819), 56 bread (224) and small amounts of flour, apples, bows, cheese and wood. Gold is 31.5% of net worth, iron 51.0%, weapons 7.7%. The town ended with 24 idle peasants and no open jobs, 14 free housing, 4 of 16 stockpile piles empty and 74 food.

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 6:00. One 25 s call at 0:00 placed three woodcutters, two hovels, two orchards and the marketplace; at 0:46 three iron mines, two dairy farms, a hovel and a chapel. High taxes at 2:05 took popularity from 100 to 70 by 5:50. The armory and its only wheat farm went up at 3:55, workshops from 4:20, a mill and two bakeries at 4:54. On the recording's count (below) net worth was under the do-nothing 1,220 from 1:00 to 9:36, low 562 at 5:30; gold was 106 at 6:00.

Iron, stone and workshops, 6:00 to 20:43. Eleven iron mines stood by 11:16 (no more sites from 13:21), four bakeries from 12:06 (no more until 26:49), three quarries from 13:57 (first stone at 19:13) and 14 workshops by 19:56. Population was 91 at 15:00 and 119 at 21:00. From 17:00 every reading shows 19 to 30 idle peasants and, but for two single readings, no open jobs. In the east oasis it put 13 woodcutters (1:39, 12:51, 16:59, 18:44) to clear trees for farms; map_overview listed 42 trees there at 0:00, 23 at 13:21 and no grove at all at 20:42.

Clearing the grove, 20:43 to 25:12. The handoff written at 20:42 lists ten woodcutters on oasis ground and says to demolish "selected" ones and rerun the farm searches, "do not demolish all until judging wood supply". The reasoning summary at 20:43 says "considering demolishing a few woodcutters could open up more farmland"; at the other nine calls it is empty or absent. It demolished all ten, six from 20:43 to 21:39 and four from 24:18 to 24:54, 5 to 10 game seconds each. Gold and wood before and after each call, from the recording:

The three gold steps (95, 95, 182) are monthly tax payments, one about every 27 game seconds; no call shows a refund in gold, and a half-cost refund on a 3-wood woodcutter is 1 or 2 wood, lost in an income of about 50 a minute. The tool's reply gives no refund figure. Each demolition cost a job: idle peasants rose from 24 to 30 by 21:50 and from 24 to 28 after the second four. No farm site opened: the wheat, orchard, dairy and hops searches at 21:18, 21:49 and 25:04 all returned "Nothing fits within 25 tiles", and no farm was placed after 3:55. Wood income fell from 54 a minute (20:43) to 12 (23:22) and 0 from 24:09, before the second four went, with 21 woodcutters standing (six rebuilt elsewhere at 21:49 and 22:47); it stayed 0 to the end with 17.

The end, 22:22 to 33:20. Extra rations (22:22), mean taxes (22:31) and extortionate taxes (24:00) gave way to full rations (25:04) and high taxes (25:12); popularity went from 90 (22:48) to 62 (28:50). Average taxes (28:56), extra rations (30:16) and mean taxes again (30:24) were its last settings. Six bakeries (26:49, 28:42) and a stockpile were its last builds, ending at 28:51; after that it traded, set taxes and rations, and waited. Population stood at 130 from 23:42 and 132 from 29:14.

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 and some other goods are left out, so it reads low against the End column, the game's figure.

Weapons

It built an armory and 14 fletcher's workshops (4:20 to 19:56) and set 13 to crossbows. The one at (403,351) stayed on bows: two set_workshop calls by tile (14:38, 14:53) opened a chapel drawn over it. The 14 made 30 crossbows and 3 bows, worth 945: 7.7% of net worth, 4.1 of the 61.1 points. From 16:31 the readout mostly gave 1.3 weapons a minute (0.3 to 2.3) for 10 to 14 workshops, about one weapon per nine workshop-minutes, against 1.0 iron a minute per mine.

The workshops cost 1,400 gold and 280 wood, the armory 5 wood, the goods 96 wood: 945 back for 1,400 gold and 381 wood, 455 below even in gold (2.0 points) or 836 with wood at its sell price of 1 (3.7). Its reasoning summaries doubted them at 17:48 ("doesn't seem economically viable for just 10 minutes of growth") and 20:43 ("isn't worth 100g"), yet it built three more at 18:44 and 19:56. None was demolished; Sol's five returned 50 gold each.

Beside its other games and the other models

Its other five games were played after this one. Its Oasis map score, the mean of its two unrounded game scores (61.1 here, 63.7 in game 2), is 62.4: third of twelve, behind Argon (76.8) and Sol (72.3). With 58.7 on the desert 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).

Third of twelve by score, net worth and population; second by gold (Sol has more) and by weapons (Argon has more); fourth by popularity; the most turns of any game here and the sixth lowest cost. It is 23.6 points above the fourth (Fable), and its 271 iron is within 15 of the two leaders', so the gap is elsewhere.

Against Sol the 4,431 gap is gold −903, bread −2,004, stone −1,330, wood −475, wheat −328, iron −115 and hops and ale −150, offset by weapons +750 and apples, flour and cheese +124. On the recording's count Terra led Sol's net worth from 9 to 21 minutes (4,333 against 4,166 at 21) and Sol led from 24 (6,675 against 5,823); Terra's population led until 18 minutes, then Sol reached 150 while Terra stood at 130 to 132 for the last 9 min 30 s. Luna has 3 mines against Terra's 11, and both ended at 24 idle.

What held it back

• No farms after 3:55. All 39 farm-site searches from 5:20 on (18 wheat, 11 orchard, 7 dairy, 3 hops) returned nothing. It had one wheat farm, one mill, five orchards and four bakeries until 26:49, and 56 bread at the end against Sol's 557.
• Wood ended. Income was 54 a minute at 20:43 and 0 from 24:09; from 26:34 it bought 105 planks for the bakeries and workshops.
• The idle cap. The rules say new peasants stop arriving while 24 stand idle. Readings sat at 19 to 30 from 17:00, with 14 to 16 free housing from 23:48; the model wrote the cap into its reasoning at 7:38 and 19:38.
• Workshops and stone. The workshops cost 2.0 to 3.7 points (above). Three quarries and 22 tethers gave 117 stone, flat from 30:38; searches at 19:38 and 20:43 found no more quarry sites; the person had 10 quarries.
• Popularity. It fell from 90 at 22:48 to 62 at 28:50 under taxes of level 6 to 8 (payments of 130 to 182 against 95 to 104 at level 5) and ended at 65.

What worked

• Iron. Eleven mines by 11:16 on every site find_sites offered; 291 iron mined (271 held, 20 sold) is 6,693 in value, and the 271 held are 51% of the final net worth.
• Taxes. About 6,000 gold came in as monthly payments, 4,400 of it after 18:00.
• Placement. 118 of 127 requested buildings were placed in 35 build_structure calls; each of the nine that were not came with a status and a remedy.
• Cache and compaction. All 157 gameplay requests read the prompt cache (98.0% of input tokens); two compactions (13:20, 20:42; 41 s together) cut the context estimate from 77,395 and 74,921 to 28,643 and 29,584.

Market

365 planks were bought for 1,460 gold in seven purchases (0:30, 3:27, 9:03, 11:51, 26:34, 28:31, 30:05), which lowered net worth by 1,095 against the sell price of 1 (9.0% of the final, 4.8 points). It sold 20 stone (140) at 3:18 and 20 iron (460) at 17:48. Against the recording all nine replies agree with the stock and gold: each plank click bought 5 for 20 gold, and the only gold differences are monthly tax payments that arrived during four of the purchases. No sale was made with 200 or more in stock (25 stone, 108 iron), so the 1.7.3 undercount did not arise.

Cost and speed

158 turns (157 answered, one failed), 264 tool calls (119 find_sites, 35 build_structure, 21 building_info, 16 each of inspect_building and set_workshop, 10 demolish_building, 9 each of market_trade, set_tax and wait_and_observe, 20 others), 7.88 million tokens and $0.727 billed (preparation $0.026, play $0.679, compaction $0.021); the interview added $0.007. Cache reads were $0.377 of the play cost. The run took 51.7 real minutes, 23.9 of them waiting for the model (median reply 8.6 s, longest 17.8 s); Sol's took 41.8, Luna's 50.3.

What 158 turns looked like. 119 of the 157 answered turns held one tool call; 37 held three to six, mostly batches of site searches (41 turns had find_sites, 26 of them nothing else). 49 of the 119 searches came back empty, 41 of them for farms. The searches and other read calls took no game time. The cost in game time was in the acting turns: 100 of them, without a screenshot, each let the host pass about 5 game seconds, 505 s in all, 25.2% of the 2,000 (Sol's 70 passes took 352 s, Luna's 61 took 321 s). Beside that, build_structure took 637 s, the nine waits 443 s, set_workshop 100, inspect_building 93, market_trade 68 and demolish_building 63. It placed 118 buildings of 127 asked in 35 calls, 5.4 game seconds per building placed. Terra's other games, by map: Oasis by the Sea with no arms placed 139 of 150 in 27 calls and 733 s (5.3); Large Barren Desert, 123 of 123 in 24 calls and 537 s with weapons counted (4.4) and 131 of 133 in 24 calls and 556 s without (4.2); The Sinkhole, 256 of 269 in 38 calls and 1,853 s with weapons counted (7.2) and 239 of 253 in 39 calls and 1,624 s without (6.8). The measure depends on map, model and host, so it is not set beside other models.

The failed request. At 23:22 a reply request failed after 20.9 s with no first token, with OpenAI's "An error occurred while processing your request. You can retry your request..." The harness logged one inference_retry (attempt 1), dropped the reply, waited 2 s and sent the request again, which answered normally. The game stayed paused (clock 23:22 before and after) and the failed request has no cost event; the delay was 22.9 real seconds and no game time. The retry list held this error before commit 55bcc41; the "terminated" case that commit added did not occur.

The interview

It put the gap down to bread: the person's much larger chain with bought wheat and automatic sales against its own "flour accumulating" and late bakery capacity, plus fear-factor efficiency and religion it had not developed. Playing again it would build the marketplace and bread industry early, test wheat purchases, add housing and jobs together, try bad-thing buildings and religion early, and watch wood sooner. It asked for clearer access to automatic trade and fear-factor effects, and said it had spent time clearing woodcutters without opening a farm site.

These match the game: the person's 59.6% more net worth, flour at 10 to 28 from 12:35 to 27:09 with four bakeries until 26:49, no bad things built, efficiency 100% against 140%, wood income that stopped, and the clearing that opened no site. Three points need a note. It would build "a marketplace" early, but it was in the first build call (0:00); the bread chain started at 4:54. "Fear-factor effects" were never looked up: its 14 distinct building_info entries include no bad thing. It does not mention the 24-idle cap, which held its population at 130 to 132 for the last 9 min 30 s with 14 houses free. No dedicated tool sets the marketplace's auto-buy or auto-sell (the rules mention both); only raw game_action clicks reach them, and it used none.

For the harness

• Farm sites. map_overview at 20:42 still listed five farmland patches (79 to 375 tiles, 1,311 in all, two of them 22 and 35 tiles from the keep), while all 39 farm searches from 5:20 on, several centred on those patches, returned "Nothing fits". The logs do not show what blocked them.
• The signpost. find_sites listed two workshop sites (448,363), (452,360) that the game refused at 3:55, and a church site (435,345) it refused at 18:06, as too close to the signpost; the church site (377,340) listed at 17:48 was "not_on_screen". The wheat farm placed at 3:55 carries the refusal's message though its status is "placed".
• set_workshop by tile. 3 of 16 calls found a chapel's panel drawn over the workshop (the third, 16:30, was redone by pixel after a screenshot, 16:51); one workshop was never switched and probably made the 3 bows.
• Default passes. 100 passes took 25% of the budget. The game saved correctly ("gpt-6-1 terra 1-7-3 8b3319").