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 79 billed requests of the run, preparation, play and compaction, went to OpenAI), reasoning medium, output limit 32,768 tokens
• Benchmark: Large Barren Desert construction (no arms), version 1.7.3 (the series records commit e0e0eb4: 1.7.3 relocked with a fix to the compaction retry that leaves play and scoring unchanged), from the save the human reference game started from; 33 min 20 s of game time (2,000 game seconds) at speed 40, close to the person's 33 min 33 s; no playbook. Weapons and armour in store count nothing toward net worth; sold at a marketplace they bring their gold, which counts.
• Run 6f469761, series 20261018T141708-5d5e0ac6, 2026-10-18: Sol's no-arms game of the 1.7.3 batch on this map, a fresh series with no memory of any other, played after the first attempt failed on an OpenRouter key limit on 2026-10-17. Its weapons-counted game on this version is run 5aad0460 (played earlier that day); its desert games from earlier rounds are run 6d57093f (1.5.6) and run 0608bbd7 (1.6.3). 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 5,145 over leaving the game alone, 107.4% of the person's 4,792, and ended 353 (3.5%) above the person's, with 105 people against 250 and popularity 85 against 100. By the project website's formula (80% net-worth share, 10% population, 10% popularity; shares run from doing nothing, 5,192 and 4 people, to the person, 9,984 and 250) the score is 98.5: 85.9 points from net worth, 4.1 from population and 8.5 from popularity. It is the best of the seven no-arms desert games listed so far (Claude Opus 5.5's 67.1 is next), 22.9 above Sol's weapons-counted game and 0.2 above its 1.6.3 game (the best score of that round). The stored goods are 570 stone (3,990), 104 iron (2,392), 20 pitch (200), 8 flour (80), 2 ale (20) and 3 wood; stone, gold and iron are 39%, 35% and 23% of net worth. The town ended with 24 idle peasants, one free place of housing and an empty granary (food −8 in the popularity factors).

How the game went

Times are the recording's game clock when a call began; the run clock in the observations reads up to 2.5 s later (median 0.8 s, 73 readings compared). A build call keeps the game running for 26 to 134 s while it clicks; the game is paused while the model replies. Its reasoning summaries are in the log (19 of 41 blocks have text).

Stone, iron and bought flour, 0 to 5 minutes. map_overview at 0:01 showed no farmland, oil or trees, stone 83 to 106 tiles west of the keep and iron 57 to 69 tiles east. Three stockpiles at 0:01, the marketplace at 0:32, three hovels at 0:53, four iron mines at 1:17 (every spot find_sites gave; a second search at 1:58 found none), four quarries at 2:14, five ox tethers at 2:59, 100 planks at 3:39, three bakeries at 4:23 and 20 flour at 4:48. Housing stood at 34 from 1:30 and the town filled it by about 4:00.

Mean taxes and a stockpile chain, 5 to 10 minutes. At 5:07, with popularity at 100, it set mean taxes (−8 a month). At 5:32 one build_structure call asked for eight stockpiles in a row toward the iron: five were placed, three refused for not touching one (134 s). The granary stood empty from 6:16 to 7:02. A warning at 7:51 said popularity had fallen from 94 to 40 (lowest reading 38); it set the bribe level at 7:52, placed an inn at 8:12 and made a second stockpile call at 8:30 (1 of 3 placed). The population fell from 34 to 29. A hovel call at 9:14 placed one of three, three more quarries followed at 9:33 and high taxes at 10:05 with the inn standing (popularity 79).

Tethers and hovels, 10 to 23 minutes. Five tethers (10:22), 100 planks (10:58), four hovels (11:37; housing 42 to 74), two more stockpile calls (12:37, 13:52: 1 of 5 placed), low taxes at 13:40 after a warning (75 to 65), and at 15:11 a mill, three bakeries and a second inn. 40 wheat (16:02), 50 bread (16:25), 100 planks (16:52), eight tethers at 17:31 and 18:28, average taxes at 19:25, five tethers (19:36), four hovels (20:50; housing 106 from about 21:30), eight tethers (21:26; 39 in all) and 20 wheat (22:58).

Waiting, 23 to 33 minutes. A 180 s wait from 23:15 ended after 178 s on a warning (popularity 88 to 76). It then built two bakeries (26:32, its last construction), set low taxes (27:02), bought 25 cheese (27:13) and waited 150 s, 100 s and 40 s; at 31:46 it set extra rations, which emptied the granary (26 food to 0) by 32:40; after the compaction it set full rations at 32:42.

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; flour, wheat, ale and pitch are left out, so it reads low (about 430 at 30:05).

Weapons

It left weapons alone. It built no armory and no weapon workshop (the final building list has none, nor does any placement in the log), made and sold no weapon, and all ten market_trade calls were purchases. None of its 19 reasoning summaries with text names a weapon or the armory; the only mentions are its own notes late on: the notebook at 30:05 ("No weapons or defenses built") and the handoff at 32:40 ("No weapons to sell"). In its weapons-counted game the same model built the armory and five fletcher's workshops (8:03, 11:24) and ended with 34 bows worth 510, for 500 gold and 173 wood. Scored by this game's rule, that game's end state is worth 8,659 (67.0 points) instead of 9,169 (75.6); that re-scores its end, it does not predict how it would have played.

Beside its other games and the other models

Against its weapons-counted game (same model, save, version and day). The gap is 1,168 as scored, 1,678 under the same rule. The two games' own files give this account; a sign is the effect on this game:

• The ground gave about the same; stone came first. Both mined iron with four mines (89 and 84, counting the 40 sold) and ended with about 105 people. This game put quarries and tethers first (2:14, 2:59; the other game built bread first, quarries at 4:12, tethers at 4:50), had stone at 8:26 against 11:25 and mined 520 against 424, on 39 tethers against 20.
• It bought less and built less with gold. Its purchases lost 2,795 of net worth against 4,145; it put 200 gold into buildings (two inns) against 1,300 (five fletchers 500, two chapels 500, three inns 300). It ran the granary near empty (0 for 274 s in 12 stretches; the other's never read 0) and ended without a bread stock, the other with 207 bread.
• Taxes ran the other way. The weapons game ended on high taxes (104 a month) and popularity 51, this one on low taxes (62) and 85. The 983 less is worth 16.4 points, the 34 more popularity 3.4.

Set the iron sale aside and, under the same rule, purchases (+1,950), gold in buildings (+1,100) and taxes (−983) make +2,067 of gold; stone (+672), the 5 more iron mined (+115), bread and flour (−1,018), ale (−160) and wood (+2) make −389 of goods: +1,678.

For Sol the weapons do not explain the moves between rounds: 47.6 with none offered, 98.3 with 260 in store, 75.6 with 510, 98.5 with none; the bows were 5.6% of the weapons game's net worth. One game of each cannot say whether the rules text changed the model's choices; its summaries here never mention weapons. Against its 1.6.3 game net worth is 48 higher, with 9 fewer people and 2 fewer inns: that game taxed at mean (136 a month), held 1,082 in bread, flour and wheat (this one 80) and set the marketplace's automatic trade (see Market). Against its 1.5.6 game it is 2,896 higher: 570 stone against 314.

Against the other no-arms desert games listed:

Sol leads Opus by 1,529: Opus held 190 more stone but kept 926 gold and ended at popularity 29. The no-arms game scored above the weapons-counted one for five of the seven (Opus +48.0, Sol +22.9, Luna +14.5, Fable +10.3, Sonnet +1.1) and below it for Haiku (−3.3) and GLM (−38.9).

What held it back

• Mean taxes at 5:07. Seven payments brought 248 gold and took popularity from 100 to 38 by 7:54; the bribe that undid it cost 144, the population fell from 34 to 29 and growth waited. No hovel was asked for between 0:53 and 9:14, with housing at 34 and 3,960 to 4,208 gold in hand from 5:00.
• The stockpile chain. Four calls (5:32, 8:30, 12:37, 13:52) took 289 game seconds, 14% of the budget, and placed 7 of 16 stockpiles; the notebook says they burned about 5 minutes. They ran east toward the iron, which read 2.7 to 3.7 a minute from 7:51 to the end; the stone, the larger value, lies west.
• An empty granary. 274 s at 0 in 12 stretches; popularity fell from 59 to 49 across 16:00 to 16:30 and from 88 to 72 across 24:30 to 27:30 as it emptied.
• Nothing built after 26:32, with 24 idle, 105 of 106 housed and 2,800 to 3,650 gold; by its notebook a quarry's first stone takes about 5 minutes.
• Extra rations at 31:46 ate the 26 food in store (104 of net worth, 1.7 points); popularity read 81 at 31:47 with +2 coming and 85 at the end.

What worked

• Stone first. 39 tethers and seven quarries mined 520 stone; stone and iron are 62% of net worth.
• Two inns against average taxes. From 19:25 it ran average taxes (−4) against two inns (+4); popularity stood at 76 to 88 from 19:30 to 25:00, and low taxes brought 62 a month at 85.
• Gold never ran short. The lowest after 3 minutes was 2,349 (23:06); the weapons game fell to 150 at about 22 minutes. No call was refused, and 13 of 17 build_structure calls placed every building asked (65 of 65).
• Warnings answered. The popularity warnings (7:51, 13:39, 26:15) got a bribe in 1 s, low taxes in 1 s and low taxes in 47 s; the food one at 22:38 got 20 wheat at 22:58. The food warnings at 5:22 and 10:21 were followed by an empty granary at 6:16 and 12:36.

Market

10 purchases, no sale: 385 planks (3:39, 10:58, 16:52, 20:18, 26:15), 20 flour (4:48), 60 wheat (16:02, 22:58), 50 bread (16:25) and 25 cheese (27:13), for 4,160 gold. Against the sell prices they lowered net worth by 2,795 (planks 1,155, wheat 900, flour 440, bread 200, cheese 100), 27% of the final value, in 224 game seconds for 108 buy clicks (2.1 s a click). It did not set the marketplace's automatic trade. The recording's input log shows each of the ten calls as the market hotkey, two selection clicks, the buy button once per lot and a right click to leave, with no drag in the game; its one game_action, a click at 22:49 on the quarry view, was outside the market, and neither its reasoning nor its interview mentions automatic trade. The 1.6.3 game did: its reasoning at 9:14 plans "automatic purchases of wood and wheat" with raw clicks, and seven game_action calls (five clicks, two drags) followed from 9:23 to 10:17. Gold here: 5,000 + 3,012 from taxes (after 144 of bribe) − 4,160 − 200 for two inns = 3,652.

Cost and speed

77 turns and 77 tool calls, 3.71 million tokens and $0.687 billed by OpenRouter (preparation $0.053, play $0.618, compaction $0.016); the interview cost $0.012 more. The agent run took 37.6 real minutes, 13.1 of them waiting for the model (longest reply 34 s). All 77 gameplay requests read the prompt cache (98.2% of input tokens).

One compaction, at 32:40, succeeded on its first request. It returned its text handoff directly (24.5 s, $0.016, 75,904 tokens, 74,558 read from the cache), with no failed request and no text-only retry, so the retry sizing that e0e0eb4 changed was not used. The last logged estimate before it was 76,192 of 120,000 tokens (31:58); the trigger is 65% of the budget. The handoff (2,785 characters) named setting rations back to full as the next step, and the next call did.

Of the 2,000 game seconds, build_structure took 784 over 17 calls, the five waits 497 (fast waits: 494.9 game seconds in 186.4 real seconds), the host's default 5 s passes after 46 acting calls without a screenshot 230, market_trade 224, place_near 85, set_tax and set_rations 66, inspect_building 35, expand_storage 25, observe 17; find_sites took none. build_structure placed 72 of the 81 buildings asked: 10.9 game seconds per building placed (7.6 without the four stockpile calls: 65 placed in 495 s). The weapons-counted game took 10.3 (672 s for 65 placed in 15 calls; 8.8 without its stockpile call), its 1.6.3 game 7.5 (534 s for 71 in 25) and its 1.5.6 game 4.3 (261 s for 61 in 28). This game and the weapons game had the host on the game machine and the two earlier ones did not, so the comparison mixes the version with the host's location.

The interview

Sol called its result a narrow net-worth win from stored stone and iron, 353 ahead of the person with fewer than half the people and less popularity: it stopped investing late and accumulated, where the person built capacity. Playing again it would make bread from imported wheat a central industry much earlier, put religion and inns first and then cut rations, grow jobs and housing together, search more widely for iron, test the fear factor and drop the stockpile experiments sooner. It found the readings and tile placement valuable; its complaint was placement reliability and unclear failure reasons, and it asked for clearer failure reasons, a read-only ledger of trade, taxes and consumption, and easier-to-anticipate action-time costs.

Checked against the run: 353 and 3.5%, more than twice the population (105 against 250), no bread sold (no sale of any kind), the 24-idle limit (from 26:15, housing 105 of 106), the placement experiments (289 s in four calls, 7 of 16 placed; five targets drew "Needs to be placed adjacent to the stockpile") and the late stop (last construction 26:32) all match. Its remark that concluding there were only four iron sites was premature fits the log: it searched for iron twice (0:01, 1:58), never again, and the person fitted seven (the interviewer's text); the 1:58 search, 13 to 15 tiles from the overview's two smaller deposits, found none within 25 tiles, and the logs do not show whether another mine would have fitted. On the fear factor, its reasoning at 1:58 and 5:07 mentions fear, and no fear building was looked up or built. What the person's settlement did comes from the interviewer's text.

For the harness

• Stockpile targets again. The sixth target of the 5:32 call was clicked with the cursor one tile east of its target (cursor tile 544 for 543) and refused; the first at 12:37 was placed the same way (543 for 542); at 8:30 the first target was displaced 5 tiles, so the next two could not touch it. The refusals' remedy names expand_storage, which it used once, at 0:01. Stopping a call at its first failed or displaced stockpile target would have saved part of the 289 s.
• Iron sites. find_sites gave four at 0:01 (five asked) and none at 1:58, while the overview lists 124 ore tiles in four deposits and the person fitted seven mines; the logs do not show whether the tool missed one.
• Quarries placed 6 tiles from the target twice (2:14, 9:33); both results say so.
• Warnings (1.7.0) fired six times: food low at 5:22, 10:21 and 22:38, popularity at 7:51, 13:39 and 26:15.
• The last call, a 30 s wait at 32:53, was answered "Session deadline" and logged as an error, the only one of 77.
• No reader watchdog collection fired; the game stayed up. The game saved at the end ("gpt-6-1 sol 1-7-3 6f4697").