GPT-6.1 Sol

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 Sol (openai/gpt-6.1-sol) on OpenRouter, served by OpenAI or Azure (no other fallbacks; all 122 requests were billed to OpenAI), reasoning medium, output limit 32,768 tokens
• Benchmark: Oasis by the Sea construction, weapons and armour in the armory counted toward net worth, version 1.7.3 (commit b967c65), from the save the human reference game started from; 33 min 20 s of game time (60,000 ticks) at speed 40, the length of the person's game; no playbook
• Run 16de155f, series 20261017T115845-938f969a, 2026-10-17: Sol's weapons-counted game of the 1.7.3 batch on this map, a fresh series with no memory of any other. Its "(no arms)" game has not been played successfully yet and is to come, so it is left out of the tables. Sol's Oasis games from earlier rounds played 1.5.1 (series 20261011T175318-657a6989) and 1.6.3 (series 20261014T024743-6703e968). 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 15,441 over leaving the game alone, 84.4% of the person's 18,293. With 150 people and popularity 74 the score is 82.3 by the project website's formula (80% net-worth share, 10% population share, 10% popularity; shares run from doing nothing, 1,220 and 4 people, to the person, 19,513 and 202): 67.5 points from net worth, 7.4 each from population and popularity. It is the highest score of the batch so far (Gemini 4 Argon, 67.4, is next) and above every Oasis score of the earlier rounds (best 68.0); Sol's two games from those rounds average 49.0. The stored goods are 276 iron (6,348), 557 bread (2,228), 307 stone (2,149), 480 wood, 41 wheat (328), weapons worth 195 and 178 in hops, ale, flour and apples. Gold is 29% of net worth, iron 38%, bread 13%, stone 13%. The town ended with 31 idle peasants, 8 empty stockpile piles and 559 food.

How the game went

Times are the recording's game clock when a call began; the run clock in the observations reads a few seconds later. A build call keeps the game running for 6 to 53 s while it clicks, and the game is paused while the model replies. Its reasoning summaries are in the log (34 of 68 blocks have text).

The opening, 0 to 5 minutes. The first stats showed a granary, so it placed none. Three woodcutters at 0:00, a wheat farm and an iron mine at 0:12. Wood stood at 1 from 0:23; the marketplace was refused for 4 wood at 1:05 and went up at 2:08. Then 55 planks (2:21; stopped short, no storage room), two stockpiles (2:36), three hovels (2:55), two iron mines (3:10), a mill and a bakery (3:38), 100 planks (4:03) and, at 4:24, one call of seven buildings (two wheat farms, three woodcutters, two iron mines; 49 s). At 5:18 it had gold 380, 28 people in housing for 34, six woodcutters and five mines.

Bread and housing, 5 to 15 minutes. At 5:18 two bakeries went up; a bakery and three hovels were refused as too close to the signpost. High taxes at 5:48 took popularity from 100 to 77 by 7:35 for 145 gold in four payments; low taxes followed at 7:35 and extra rations at 7:44. Food in store was 11 at 7:35, touched 0 at about 8:45 and was 45 at 9:24 (popularity bottomed at 72 at 8:56). Meanwhile came two quarries and three tethers (6:31; gold 105 after 75 planks), five bakeries (7:00; eight in all) and two more quarries (7:52). It sold 20 iron for 460 at 9:24 to buy planks, placed five iron mines at 9:55 (ten), then three hovels, a mill, two bakeries and a wheat farm (10:59), two hops farms, three breweries, an inn and two bakeries (12:16) and five bakeries (12:57; 17 in all). Popularity was 95 or more from 14:03.

Taxes, iron and the last sites, 15 to 23 minutes. At 15:24 it sold 70 stone for 490, then set double rations (15:40, +8 popularity) and mean taxes (15:48, −8; 108 gold a month with 90 people). Three iron mines and five hovels (16:17) made 13 mines and housing 130, five bakeries and three tethers (17:02) made 22 bakeries, and at 17:40 find_sites found no iron site within 25 tiles. It read the fletcher's and poleturner's entries (17:40), bought 20 hops and 10 ale for 500 gold (17:54, 18:05) and at 19:58 placed three inns, three fletcher's workshops and an armory in one call (44 s). The handoff written at the compaction (20:47) shows 22 idle, popularity 100 (double rations +8, mean taxes −8) and a plan to raise taxes once the inns gave a bonus. Searches at 22:20 found no site for iron, quarries, wheat farms or pitch rigs. Taxes went to extortionate at 22:42 (182 gold a month); two more fletchers and a hovel came at 22:50, two granaries at 23:17.

Jobs for the idle, 23 to 29 minutes. A 120 s wait from 23:31 took gold from 1,329 to 2,077 with 134 people in housing for 138 and 24 idle. Then full rations (25:32), high taxes (25:40), 30 wheat for 690 gold (25:49) and three stockpiles (26:02). Three batches of woodcutters (7 at 26:17, 5 at 27:01, 4 at 28:06; 16 in all, two with a hovel) were followed by a rise in population from 134 to 150 by 28:44. Mean taxes returned at 27:58.

Waiting and demolishing, 29 to 33 minutes. Its last construction call ended at 28:40. After a 120 s wait from 28:44 it demolished five fletcher's workshops (30:45 to 31:35) and two inns (32:49, 33:01), 50 gold back for each, and waited out the rest; the last call, a 10 s wait, returned "Session deadline".

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, wheat, flour, hops and ale are left out, so it reads low (about 660 in those goods at the 28:44 reading).

Weapons

It built an armory (5 wood) and three fletcher's workshops at 19:58, set them to crossbows (3 wood, 30 at the sell price) at 20:48 to 21:29, and added two fletchers at 22:50 that were never set (new workshops make bows: 2 wood, 15). It built no poleturner, blacksmith, armourer or tanner. The first weapon reached the armory between 23:30 and 25:32, 3 to 5 min after the workshops were placed; by the end the five had made 5 crossbows and 3 bows, worth 195: 1.2% of net worth, 0.9 of the 82.3 points. The readout put the five together at 0.3 to 1.3 a minute.

The workshops cost 500 gold and 100 wood (100 gold and 20 wood each, prompt/game-mechanics.md), the armory 5 wood and the goods 21 wood (5 × 3 and 3 × 2): 500 gold and 126 wood. Demolishing the five at the end returned 250 gold (a 50-gold jump in the recording at each of the seven demolitions). So 195 in weapons and 250 in gold came back for 500 gold and 126 wood: about 181 below even, 0.8 points (1.9 before the refund). Its reasoning at 17:40 weighed a metal-armour workshop, found it poor and preferred wooden weapons; the handoff says the fletchers were built "to turn excess wood into stored value and jobs".

What else the town could have done: the 500 gold kept as gold was worth 2.2 points. An iron mine (20 wood, no gold) delivered about one iron a minute, 23 in value, against about 16 a minute for the five workshops together (195 over 12 min), but iron sites were gone by 17:40 and quarry sites by 22:20. Woodcutters (3 wood each) were what it built last, and 16 more people followed. Weapons do not explain the move from Sol's 1.6.3 game (4 fletchers, 345 in weapons): they fell by 150 while net worth rose by 6,208.

Beside its other games and the other models

How it reached 16,661. Against its own 1.6.3 game the gain of 6,208 is gold +2,063, bread +2,148 (557 against 20), iron +2,001 (276 against 189), hops, ale, wheat and apples +278, stone −42, flour −90 and weapons −150. Three things in the logs lie behind it:

• A bread chain that fed 150 people. The first mill and bakery went up at 3:38; there were 8 bakeries by 7:31, 17 by 13:18 and 22 by 17:35, with 2 mills and 4 wheat farms. The population was 89 at 15 min and 129 at 21 min; the other six weapons games ended with 50 to 84 (Gemini 4 Argon, 167).
• Taxes held up by that food. The recording shows 61 monthly payments from 6:09, about 6,800 in sum; the gold accounting (4,755, less 1,000 at the start, 950 from sales and 350 from refunds, plus 3,710 spent in the market and 900 on inns and fletchers) gives about 7,000, about 6,000 of it after 15:48. Double rations (+8) offset mean taxes (−8), and the inns' +6 from 21:29 offset more, so popularity stayed at 95 or above from 14:03 to 29:02 while it paid up to 182 gold a month. Gemini 4 Argon also taxed (166 a month); the other six games ended with no taxes or at most 84 a month.
• The same land as the best others. 13 iron mines (the person had 13) gave 296 iron (276 held, 20 sold), against 223 to 246 held in the Sonnet, Opus and Fable games; 5 quarries left 307 stone in store after 70 were sold, against 25 to 161 in the Claude games. Its net worth at 15 min (2,202) matched Sonnet's (2,205); the gap opened later (6,675 against 4,513 at 24 min).

Sol's weapons were 1.2% of its net worth, lower than every other game's but Fable's (0.6%); Gemini, with the most (1,272), had 9.7%. The two top scores are the two games with the most people and taxes above 160 a month.

The 1.5.1 game also built a large bread chain (16 bakeries) and held popularity at 100, with 130 people. In the 1.5.1 rules weapons added nothing and Sol built none; in 1.6.3 it built four fletchers, here five.

What held it back

• The cheap sites ran out. All 13 iron mines stood by 16:57; at 17:40 and again at 22:20 find_sites found no mine spot, and the second stone deposit (349, 396) gave no quarry spot either time. Stone first arrived at 12:38, 6 min after the first quarries (6:31); the person had 10 quarries.
• Population stopped behind jobs. Idle peasants stood at 20 to 24 in every reading from 19:05 to 30:45. The rules say new peasants stop arriving while 24 stand idle; the handoff noted it and chose net worth over "useless job spam". Population sat at 134 (housing 138) from 23:30 to 26:17 and rose only after the woodcutter batches. Nothing was built in the last 4 min 40 s, with 22 to 31 idle and housing 154 for 150.
• Gold into side bets. Four inns (400 gold, 80 wood), five fletchers (500 gold, 100 wood) and 20 hops with 10 ale (500 gold, 240 of net worth lost) are 1,400 gold. The inn factor read 0 from 12:57 to 21:00, +2 at 21:15 and +6 at 21:29, once ale came; ale in store fell from 11 to 1 by 23:17.
• Planks. 630 planks cost 2,520 gold and 1,890 of net worth (11% of the final). It had six woodcutters from 5:13 to 26:17 and bought planks seven times, the last at 18:30.
• Thirty wheat for 690 gold (25:49) lost 450 of net worth and was not used up: wheat in store went from 2 to 31 and 41 at the end while flour read 1 to 5 in the last 2.6 min, which points to the two mills, not wheat, as the bakeries' limit.
• The tax raise at 27:58. Mean taxes (high before) added about 31 gold a payment (180 against about 150 for 150 people, at the 1.0 a person it paid under high taxes), about 370 in 12 payments or 1.6 points. The popularity drift went from 0 to −2 a month, taking popularity from 99 to about 76 (2.3 points), so the raise lost about 0.7 points net. The two inn demolitions then took about 2 more popularity points (0.2) for 100 gold (0.4).

What worked

• The bread chain from 3:38. After the dip at 8:45, food in store stayed at 23 or more in the recording and ended at 559; 557 bread are worth 2,228.
• Taxes with rations and inns: popularity of 95 or more for 15 minutes at 108 to 182 gold a month.
• Thirteen iron mines by 16:57, every site find_sites offered; 20 iron sold at 9:24 paid for planks without lowering net worth.
• Woodcutter batches at 26:17 to 28:06 and the 16 extra people that followed.
• Batching and placement. 22 build_structure calls placed 111 of 118 buildings asked (up to 8 a call). Every building reported as placed is in the final count (less the five fletchers and two inns demolished). No call was refused; seven placements did not go through (below).
• The prompt cache held on all 120 gameplay requests (98.3% of input tokens).

Market

630 planks (2,520 gold), 20 hops (300), 10 ale (200) and 30 wheat (690) were bought for 3,710 gold, which lowered net worth by 2,580 against the sell prices (planks 1,890, wheat 450, hops 140, ale 100), 15.5% of the final 16,661. It sold 20 iron (460) at 9:24 and 70 stone (490) at 15:24; selling leaves net worth unchanged. The seven plank purchases (2:21, 4:03, 6:13, 9:35, 11:55, 15:56, 18:30) took 106 game seconds, all 12 market calls 143. These are units at the market's fixed prices.

Cost and speed

120 turns, 125 tool calls (35 find_sites, 22 build_structure, 12 market_trade, 8 waits, 7 inspect_building, 7 demolish_building, 6 place_near, 6 set_tax, 5 building_info, 4 expand_storage, 3 each of update_plan, set_rations and set_workshop, 2 each of map_overview and center_on), 5.65 million tokens and $0.970 billed by OpenRouter (preparation $0.053, play $0.892, compaction $0.026). The context was compacted once (44 s, at about 20:47) and ended at about 62,000 of its 120,000 budget. The interview cost $0.016 more, with 60,715 of its 62,300 tokens read from the cache. The agent run took 41.8 real minutes, 16.7 of them waiting for the model. All 120 requests read the cache (98.3% of input tokens).

Of the 2,000 game seconds, build_structure took 667 over 22 calls, the 8 waits 568, the host's default 5 s passes after 70 acting calls without a screenshot 352, market_trade 143, inspect_building 58, demolish_building 53, place_near 52, expand_storage 43, set_tax and set_rations 31 and set_workshop 25; find_sites took none. Five of the eight waits ran fast-forwarded (480 game seconds in 167 real seconds). build_structure placed 111 buildings (118 asked), 5.0 a call: 6.0 game seconds per building placed. Its 1.6.3 game took 7.2 (82 placed) and its 1.5.1 game 2.8 (74). This game had the host on the game machine; none of Sol's earlier games did (the first hosted there were Fable's, on 2026-10-16), so that comparison mixes the version and the host's location. On this map in the batch the measure ran from about 6 s a building (Sol, Gemini 4 Argon, GPT-6 Luna at 7.4) to 14 to 20 s (the Claude models): placing about five buildings a call, as this game did, keeps it low.

The interview

Sol put its result at about 85% of the person's net worth with fewer people and lower popularity, and said the gap was investment allocation and timing: the person ran a larger, faster engine (57 bakeries against its 22, twice the quarries, 140% worker efficiency) while it spread money over food, mining, ale, taxes and weapons. It called the inns and fletchers disappointing, said it accepted slow stone for too long, that its lower popularity was partly deliberate for gold, and that it demolished workshops and two inns near the end to recover costs. Playing again it would build bread, wheat purchases and worker efficiency earlier, add housing and jobs together, revisit quarries and ox transport, and skip ale, inns and fletchers. It found the tile placement, storage and production readings helpful and wanted goods waiting at producers, consistent signpost checks and a clearer account of batching and time costs.

Against the game: 85% (85.4%), 22 bakeries, 5 quarries against 10, 100% efficiency, buying wheat late (25:49), reaching the idle cap with housing free (23:30 to 26:17), the signpost refusals (a bakery at 5:18 and a brewery at 21:41 were listed by find_sites and then refused), the inns and fletchers, the deliberate popularity decline and the end demolitions (five fletchers, two inns, 350 gold) all match. Two statements need a note. "Not that the person held more gold" holds only for the sum: the person held 13,549 more gold and 10,697 less in goods, a gap of 2,852. And the demolitions added 350 gold and some wood but the two inns cost about 2 popularity points, roughly 0.2 points net. Its remarks on religion and fear buildings are not tested here; it built none.

For the harness

• The granary. The rules' Start paragraph tells the model to place the required granary, but this save has one. Sol read "granaries: 1" in its first stats and placed none, so no first call was rejected.
• The signpost. find_sites offered a bakery spot (451, 363) and a brewery spot (431, 346) that build_structure then refused as too close to the signpost (5:18, 21:41); three hovels it chose by tile at (432 to 442, 348) were refused the same way at 5:18. These are 5 of the 7 buildings that did not go through; the others were an ox tether (no spot within 4 tiles, 6:31) and a woodcutter (no open side, 28:06). A mill by place_near (3:26) and two of three hovels (10:23) also found no free spot.
• The production readout. From 17:39 it credited the 22 bakeries with 56 to 92 bread a minute while the two mills delivered 9 to 11 flour a minute; the handoff takes bread as 72.3 a minute. Bread in store rose by 34 a minute over the last 12.4 min (135 at 20:48 to 562 at 33:13). I did not reconcile the readout's figures.
• Demolition refunds. Each of the seven demolitions returned 50 gold, but the tool's result does not say so; the model's reasoning summary at 31:00 speaks of "a refund of 50 wood and 10 resources".
• Default passes. After 70 acting calls the host let 5 game seconds pass (352 s, 17.6% of the budget), which the interview asks to have explained more clearly.
• The game saved correctly at the end ("gpt-6-1 sol 1-7-3 16de15").