Kimi K3.1 Pro

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: Kimi K3.1 Pro, through Moonshot's pre-release endpoint for it, kimi-1021-preview (released as kimi-k3.1-pro), reasoning effort high, output limit 32,768 tokens; costs at Kimi K3.1 Pro's list prices ($2 per million input tokens, $10 output, $0.20 cached, $2 cache write), as Moonshot reports no cost per request.
• Benchmark: The Sinkhole construction, version 1.7.1 (commit 07f2b7a), that benchmark's round since 2026-10-15, from the save "The Sinkhole start"; 45 min of game time at speed 40, the length of the person's game; no playbook. Weapons and armour in the armory count at the rules' prices. The host ran on the game machine. The runner checked out 07f2b7a for it, like the other models' listed games here; the model's other five games ran on 1.7.3 (commit 55bcc41). 1.7.1 has no retry for a reply cut off by the provider closing the connection; none was needed (70 requests completed, none failed).
• Run 3d6c968e, series 20261023T184800-d00fe445, 2026-10-23: game 5 of this model's six games that day, each a fresh series with no memory of any other. This game counts on the website's leaderboard.

Result

A full-budget score: the run used all 2,700 game seconds. Net worth grew by 11,678 over leaving the game alone, 29.6% of the person's 39,436. By the website's formula (100 × (0.8 × net-worth share + 0.1 × population share + 0.1 × popularity ÷ 100), shares from leaving the game alone (net worth 2,195, population 4) to the person (41,631, 506), net-worth share clipped to −1..2) the score is 31.78: 23.69 from net worth, 2.59 from 134 people, 5.50 from popularity 55. With this game added it is sixth of the eleven 1.7.1 Sinkhole games with weapons counted.

The stored goods are 202 iron (4,646), 92 stone (644), 33 spears and 7 bows (435), 34 apples (136), 35 wood and 2 cheese: gold is 57% of net worth, iron 33%. The town ended with 134 people in 146 places, 24 idle peasants, no open job, extra rations and 36 food.

How the game went

Times are game time; for a call, when it began (the recording's clock; the run clock in the observations reads a few seconds later). Taxes start at none.

0:00 to 6:11. One build_structure call (87 s) put up the marketplace, a hovel, three woodcutters and two orchards: "Oops building call took 88 game seconds!" It bought 100 planks at 1:34 (54 s), put up an armory, two fletcher's workshops, a poleturner's, two hovels and an orchard at 2:34, and two iron mines with a fletcher's, two poleturner's, two hovels and an orchard at 4:10.

6:11 to 18:43. Average taxes at 6:11. The meat ran out and one food kind was eaten: food read 12 for 47 people at 9:22 and popularity fell from 100 to 60 by 11:29. It set low taxes (11:30) and none (13:29, popularity 51; the recording's lowest is 50 at 13:31). Two chapels (7:42, 14:19; 250 gold each) and five poleturner's workshops (12:05) were paid from gold, which fell to 182 at 14:19, its lowest, and stayed there until the first weapon sale at 16:04. Then came extra rations (16:59; the reply: +4 popularity), average taxes (17:12), four more poleturner's (17:28) and high taxes at 18:43.

18:43 to 32:30. No workshop was built after the 17:28 call. It turned to iron mines (3 at 19:54, 4 at 22:08, 3 at 28:42), with a quarry and two ox tethers at 19:54, orchards, dairy farms and hovels. One compaction came at 24:55. Population passed 100 at 30:05; high taxes from 28:30.

32:30 to 45:00. It sold 35 spears (32:29), bought 100 planks (33:00) and put up 6 mines (33:57), 3 more (36:02) and a hovel and five orchards (38:24, a 159 s call). Food read 0 three times in the recording (40:27, 40:33, 40:46), inside that call. Population reached 134 at 41:03. It set full rations at 41:52 (food 11) and extra again at 43:47, and waited 60 s (42:41) and 45 s (44:12) at the end; popularity went from 70 at 41:08 to 55.

Idle peasants reached the cap of 24 at 9:22 and read 18 to 24 after. Open jobs read 0 in every observation but one (1 at 13:27). Popularity never fell under 50, and population never fell by more than 1 at a time.

Spells run from each set_tax call to the next; they sum to 6,469 gold in 79 paydays. The one payday under "none" (25 gold, 13:38) came two seconds after the reply.

Each column is the last reading at or before that minute. Net worth adds the weapons from the last observation before the minute (up to 97 s earlier). Its lowest reading was about 1,070 at 9:10 (gold 605), and it passed its start of 2,415 again at about 22:00.

Weapons

At 2:34 it built an armory and the first workshops (at 1:34 it had reasoned that a bow turns 2 wood into 15 of value) and added to them at 4:10, 12:05 and 17:28: 3 fletcher's and 12 poleturner's workshops, 1,500 gold and 180 wood, plus 5 wood for the armory. set_workshop was never called (fletchers made bows, poleturners spears).

They made 262 spears and 61 bows, worth 3,535 at the rules' prices (spears 10, bows 15). 3,100 of it was sold in 12 sales (229 spears, 54 bows) and 435 is held (33 spears, 7 bows): 3.1% of net worth and 0.9 of the 31.8 points (30.9 with weapons worth nothing). The pieces took about 384 wood (1 per spear, 2 per bow), so at the rules' prices, wood at 1, the wing returned 1,466 over its costs, about 3 points. Held plus sold, weapons come to 3,535 here, the most of the eleven games; next are GLM 5.3 Flash (2,265) and Claude Opus 5.5 (1,530).

At the 43:46 reading the 12 poleturner's delivered 8.3 spears a minute (0.7 each) and the 3 fletcher's 1.3 bows (0.4 each); in the same readout the 21 mines delivered 0.7 iron a minute each, an orchard 1.7 apples and a dairy farm 1.6 cheese. At sell prices that is about 16 gold a minute for a mine, 7 for a poleturner's, 7 for an orchard and 6 for a dairy farm. At 18:43 its own estimate was the same: poleturner's "payback18min, not great!", mines "Much better." It built no more workshops.

The armory holds 50. expand_storage (11:44) was refused: "no free ground touches it on any side", with a remedy to sell weapons or demolish a neighbour. The model read it as its own layout (workshops placed round the armory) and decided "Selling output doesn't reduce net worth, so simplest sell when full." The armory read 48 of 50 at 16:03 and 50 of 50 at 32:28 and 41:23, each followed by a sale within seconds; the 3-minute spear rate fell from 7 a minute (30:24) to 4.0 (32:58) and was 7 again by 35:14. After 25:00, selling took 157 game seconds.

Beside its other games and the other models

Scores use each map's own scale. On this save the no-arms game, run on 1.7.3, ended with 4,420 more net worth, 35 more people and 19 more points of popularity.

Ordered by score. "Sold for" is the gold in the replies of weapon and armour sales; the map score is the mean of the model's two Sinkhole games (weapons counted and no arms), from unrounded game scores. This game is sixth by score and fifth by net worth; it has the fourth most people, the seventh best popularity and the third most gold (after Gemini and Sol). Its map score is third of eleven, about 0.6 above Claude Opus 5.5; the no-arms game (43.3) is third in its benchmark. On the other maps its map scores are 40.9 on Oasis (third) and 17.8 on the desert (seventh); the mean of the three maps is 32.1, fourth of eleven, level with Claude Sonnet 5.5.

Kimi K3 on the same benchmark. The earlier model played it on 2026-10-15 (run ff80a5af, series 20261015T175751-2419d558, not on the website): version 1.7.1 at commit 674e7ba, a later relock of the same version (two harness changes whose commit messages each say what the model plays is unchanged), reasoning high, output limit 32,768, the same prompt and save.

The new model ended with 1.9 times the net worth and score, placed almost twice the buildings, waited a twelfth as long and put 21 mines on the iron where K3 put 5.

What held it back

• Jobs, not housing. The town was 110 at work and 24 idle: arrivals stop while 24 stand idle, and 12 places stood free at the end. It knew ("Pop65, idle24 cap prevents new arrivals despite housing98", 16:59) and built orchards and mines for their jobs as well as their output, sizing each call to the wood on hand and the time a call takes (its reasoning at 28:05 and 37:41).
• Placement took most of the clock. build_structure took 1,660 of the 2,700 game seconds (61.5%): 15.7 s per building placed, 70 to 159 s a call for 6 to 8 placements. The game runs meanwhile; the result was 106 buildings.
• Popularity swings. It fell from 100 to 51 in the first 13 minutes (average taxes, one food kind eaten at times, religion 0). The two chapels, 500 gold, showed religion +2 in 8 of the 47 readings that list it (18:41 to 18:56, 28:04 to 30:24, 41:51 to 42:03). It fell 15 points in the last four minutes, 1.5 of the 31.8: under full rations (41:52 to 43:47) the +4 was gone, religion went back to 0 and taxes stood at −6 or −4.
• Iron sites ran out. The nearby mine sites were all used by 37:41 ("Nearby mines all used"). Iron delivered per five minutes was 0, 2, 7, 8, 10, 25, 31, 51, 68: mines take minutes to reach 0.7 a minute (18 stood at 35:14 delivering 7.0 a minute, 21 at 41:08 delivering 13.3).
• The gold (see Market): 7,969, 57% of net worth, sat from 29:38 on.

What worked

• Iron. 21 mines for 420 wood mined 202 iron (4,646), none sold: 9.4 of the 31.8 points before the 420 wood.
• Taxes with extra rations. 6,469 gold in 79 paydays; extra rations (+4 a month from 16:59) let it run average and high taxes with popularity staying at 50 or more, and no peasants left.
• Selling weapons. 12 sales kept the armory working and put 3,100 gold in hand; the wing returned 1,466 over its costs.
• Placement. 106 of 113 placements were made. Of the 7 that were not, one was 2 wood short, two came back unverified and four not_placed, each with a remedy. It retried only the hovel (placed at 9:24) and the (380, 312) orchard (asked again at 38:24, refused again).

Market

16 calls, 449 game seconds (16.6% of the budget). Purchases: planks 400 (1,600 gold) in four calls of 100 (20 clicks, 50 to 54 s each), bought at 4 and counted at 1: 1,200 of net worth lost. Sales: spears 229 (2,290) and bows 54 (810), 3,100 in 12 calls, leaving net worth unchanged.

The gold. It reconciles: 2,000 start + 6,469 paydays (taxes) + 3,100 sales − 1,600 purchases − 2,000 buildings (3 fletcher's 300, 12 poleturner's 1,200, 2 chapels 500) = 7,969. So 6,469 came from taxes and 3,100 from weapons. After the last workshop (18:07) the only gold spent was the 400 for planks at 33:00; mines, farms and hovels cost wood. Gold passed 3,000 at 29:38, 4,000 at 32:47, 5,000 at 36:44, 6,000 at 39:25 and 7,000 at 42:05. At 37:41 it put a new workshop's payback at 35 minutes, and at 41:09 wrote "Could convert excess gold to goods? Buying always lowers net worth, no." Holding gold costs nothing in the score; the rules' gold-priced options were planks (bought at 4, counted at 1), workshops and chapels.

The recording agrees with all 16 replies. The 12 sales show the clicks the replies count (5 spears or bows at 50 or 75 gold a click, the last click short in two: 4 bows at 16:35, 4 spears at 21:36) and the four purchases 20 clicks of −20. Where a call's gold change differs from the reply's goldChange, the difference is a payday inside the call: 12 paydays in 10 calls, 1,068 gold. No stock was near 200.

Cost and speed

70 turns, 127 tool calls (63 find_sites, 16 build_structure, 16 market_trade, 11 set_tax), 3.67 million tokens and $1.287 at Kimi K3.1 Pro's list prices (preparation $0.033, play $1.219, compaction $0.035); the interview cost about $0.037 more. Moonshot reports no cost. The run took 54.0 real minutes, 15.4 of them waiting for the model; one compaction came at 24:55 (33 s) and one fast wait covered 60 game seconds in 27.7 real ones. All 70 gameplay requests read the prompt cache, 96.9% of input.

Of the 2,700 game seconds, build_structure took 1,660 (61.5%), market_trade 449 (206 buying planks, 243 selling weapons) and waits 110. That is 15.7 game seconds per building placed by build_structure (106), against 13.6 (Oasis, weapons; 80), 12.9 (Oasis, no arms; 93), 12.0 (desert, weapons; 48), 12.7 (desert, no arms; 67) and 14.9 (Sinkhole, no arms; 129) in the model's other games, all with the host on the game machine; the maps differ, so this says nothing general.

The interview

It put the gap down to scale: 134 people against the person's 506, and a popularity-led economy (inns, ale, rations cut, a bread chain, religion) against its own farms and wooden weapons. Playing again it would build inns and buy ale first, cut rations, build wheat and bread, lift the idle cap with jobs, put its effort into iron and quarries rather than fletchers and poleturners, and batch construction by area. It called the benchmark "clear and fair" and asked for bulk trading, same-area batch construction, verifiable placement outcomes and one reference for the economics. Against the game, "roughly three times my net worth and nearly four times my population" (3.0 and 3.8), the plank purchases and off-screen placements that "could consume substantial game time" (206 s for four purchases; calls of 70 to 159 s), the sites that "failed or returned unverified" (7 of 113) and the failed armory expansion match. These do not, or only partly:

• Extra rations "limited popularity": they added 4 points a month (the set_rations reply; a food factor of +4 to +5 in the readings from 17:10). They cost food, which read 0 three times, and popularity fell after its switch to full rations at 41:52.
• Workshops "had relatively poor returns": partly. They returned 1,466 over their costs at the rules' prices, and per building they earned less than a mine (7 against 16 gold a minute), as its 18:43 estimate said.
• A late game spent "reacting to full armories, fluctuating religion, and low food": after 25:00, building took 681 game seconds and weapon sales 157; the armory read full twice.

For the harness

• Placement on 1.7.1. Four not_placed replies carried retrySkipped: no_free_spot_nearby and the remedy to use find_sites (dairy farm 25:40, iron mine 28:42, orchards 30:25 and 38:24). The mine was a site listed at 18:58, overlapping a mine placed at 22:08. The orchard at (379, 312) / (380, 312) was listed by find_sites at 30:25, 32:29 and 37:41 and refused at 30:25 and 38:24. A hovel in the 6:27 call was not_placed 2 wood short, with the remedy to wait or buy. Two placements came back unverified (quarry 19:54, orchard 36:02: "Check the next observation or status before placing it again"); the standing counts showed neither was built and it did not place them again. 1.7.1's retry searched up to 12 tiles: at 38:24 an orchard asked at (309, 345) was placed at (298, 334), 11 tiles away on each axis (offsetTiles [−11, −11]), and at 2:34 a fletcher's workshop moved one tile. The model's later reasoning does not mention the move. 1.7.2 limits that search; this game cannot say what that would have changed.
• expand_storage on the armory. The reply (11:44) said no free ground touched it, and the model never retried; it sold 283 weapons instead. 1.7.2's room-keeping rule covers armories placed flush with their kind; nothing here shows whether it would have left a side free.
• Warnings. Six fired (popularity falls at 8:25 and 11:29, low food at 9:22 and 22:06, low gold at 13:27 and 19:52) and none after 22:06. The 159 s placement call at 38:24 had no observation inside it; the recording shows food at 0 in it, and the 41:08 observation carried no warning, with 22 food for 134 people on extra rations.
• The game saved at the end ("kimi k3-1 preview 1-7-1 3d6c96", 3,752,667 bytes in episode.json).