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. All 71 recorded requests (1 preparation, 69 gameplay, 1 compaction) succeeded.
• Benchmark: Oasis by the Sea construction, version 1.7.3 (commit 55bcc41, 1.7.3 relocked: it adds a retry for a model reply cut off by the provider closing the connection; what the model plays is unchanged, and no inference_retry event fired in this game), from the save the human reference game started from; 33 min 20 s of game time (60,000 ticks) at speed 40; no playbook. Weapons and armour in the armory count toward net worth at the rules' prices. The host ran on the game machine.
• Run 0a155c24, series 20261023T154701-a8c5da72, 2026-10-23: game 1 of Kimi K3.1 Pro's six ranked games, a fresh series with no memory of any other. It is the third start of this benchmark: two earlier attempts in series 20261023T145826-2c77215a (runs e97c00a2 and 5da919d8, on commit 0a58665 without the retry) were cut off when Moonshot's server closed the connection mid-reply, at about 18 and 11 game minutes, which the runner counts as an infrastructure failure and does not score. 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 6,215 over leaving the game alone, 34.0% of the person's 18,293. By the website's formula (80% net worth, 10% each population and popularity, the net-worth share clipped to −1..2) the score is 37.3: 27.2 points from net worth, 5.2 from population (102 of the person's 198 above the 4 who live there) and 5.0 from popularity.

The stored goods are 143 iron (3,289), 152 stone (1,064), 575 wood (575), 7 crossbows (210), 8 pikes (144), 31 bread (124) and small amounts of apples, bows, spears and cheese. Iron is 44% of net worth, iron and stone 59%. The town ended with no free housing, 23 idle peasants, no open jobs, 1 of 20 stockpile piles empty and 37 food.

How the game went

Times are the game clock when a call began (the recording's clock; the run clock in the observations, which supplies the idle, popularity and production readings quoted here, reads a few seconds later). The game is paused while the model replies; a build or buy call keeps it running while it clicks (up to 118 s here).

The opening, 0:00 to 3:11. After the map overview, eight building entries and five site searches (all free), one call of 108 s placed two woodcutters, the marketplace, an iron mine, two hovels and an orchard. Its reasoning afterwards: "Oops build consumed 1m51 game!" At 1:56 it bought 100 planks (400 gold; wood 2 to 102) and at 2:41 added two stockpile piles. Population reached 26 of 26 housing, with 21 idle, by 3:25.

Taxes and housing, 3:11 to 9:25. High taxes at 3:11 (26 gold a month) took popularity from 100 to 56 by 7:01; taxes went to none at 6:44. The 3:24 call (two mines, a quarry, a tether) lost a hovel ("Too close to signpost to build") and the second quarry; the 4:40 call (118 s) placed the armory, a fletcher, a poleturner, a dairy farm, a woodcutter and an orchard, but not the chapel or the hovel. Population sat at its housing from 3:11 to 7:56. Food fell to 7 by 9:26; it sold 33 stone (231 gold) at 8:59 and bought 40 bread at 9:25.

Mines and quarries, 9:53 to 17:21. Low taxes at 9:53, 50 planks at 10:06, then three mines, a quarry, an orchard and two hovels (10:38), two mines (12:51) and the ninth mine (14:18); average taxes from 12:40. Gold hit 72 at 14:30.

Woodcutters, food and the idle cap, 17:50 to 30:04. The wood readout fell to 8 a minute at 17:51 (16 at 17:22); it put four woodcutters at the big grove at 17:50 and five at 19:00. The one compaction came at 19:56; four hovels at 20:27 raised housing from 74 to 106. Orchard searches found nothing at 21:31, 21:56 or 22:19 ("No farms anywhere, must clear or buy"). A food warning (21 food for 81 people) came at 22:19; it sold 40 iron (+920) and bought 100 bread (−800) at 22:42. It then had 24 idle and 82 people in 106 housing, and reasoned that peasants stop arriving while 24 are idle. Eight woodcutters at each of 23:36, 24:46 and 28:18 gave jobs; a fourth chapel came at 26:09, high taxes at 26:36.

The last three minutes. It sold wood at 30:11, bought 20 cheese at 30:37, waited 60 s at 31:01, set mean taxes at 32:06 and average at 32:54 as food variety lapsed. Popularity went from 55 to 50.

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 are left out, so it reads low. On that count it was under the do-nothing 1,220 from 2:16 to 15:07 (low 722 at 10:32) and under its own start of 1,380 until 16:00.

Weapons

It built an armory, one fletcher's workshop and one poleturner's workshop in the 4:40 call and set them to crossbows and pikes at 7:00 and 7:21 (its reasoning put crossbows at 10 gold per wood, pikes at 9). The readouts gave 0.3 to 0.7 a minute whenever a workshop delivered. At 9:53 it wrote "Production workshops very slow 0.3/min each", at 13:33 "Prioritize mines", at 19:00 "Workshops not worth late." In 28 minutes the fletcher made 8 (1 bow, 7 crossbows) and the poleturner 9 (1 spear, 8 pikes).

At the end the armory held 379: 1 bow (15), 7 crossbows (210), 1 spear (10), 8 pikes (144). That is 5.1% of net worth and 1.7 of the 37.3 points (35.7 with weapons worth nothing). Nothing was sold. From the rules' table the two workshops cost 200 gold and 30 wood, the armory 5 wood, and the weapons used 40 wood (2 + 21 + 1

• 16): 275 at store value, so the weapons ended about 104 ahead (0.45 of a point). Per building a fletcher made about 8 gold of weapons a minute and a poleturner 5, against about 27 for an iron mine (1.2 iron a minute at 23).

Beside its other games and the other models

Ordered by score; Kimi's cost is at its list prices. This game is fourth of eleven by score (Fable is 0.2 above), fifth by net worth, third by population, level with Sonnet and Muse on popularity, and sixth by weapons. Its 143 iron compares with 223 to 246 at Sonnet, Opus and Fable; the 80 to 103 it lacks is worth 1,840 to 2,370. It has the most woodcutters in this table.

Its Oasis map score, the mean of this game and the no-arms game, is 40.9 (37.3 and 44.5), third of eleven: Gemini 76.8, Sol 72.3, Kimi 40.9, Fable 39.0, Muse 38.0, Opus 35.3, Sonnet 33.9, Luna 33.8, DeepSeek 27.2, Haiku 25.0, GLM 21.1 (each from its two games' unrounded scores). Its desert mean is 17.8 (seventh) and its Sinkhole mean 37.6 (third).

What held it back

• Little wood, then too much. Four woodcutters delivered 3.3 to 6 wood a minute each until 17:22 (wood 2 at 1:57, 3 at 15:58); it bought 195 planks for 780 gold. It then added 33 woodcutters in 12 minutes, 37 in all, 46% of the 80 buildings it placed. At 33:11 they delivered 72 wood a minute (about 1.9 each) against 10.7 iron from nine mines (about 27 each). 575 wood stood unspent, 7.7% of net worth.
• Nine mines, four quarries. Three mines stood by 4:33, six by 12:10, nine by 15:14; searches then found one site (13:33) and none (21:31). Sol, Gemini, Fable and DeepSeek fit 11 to 13 on the same ore. The person had 13 mines, 10 quarries and 33 tethers (here 9, 4 and 5).
• Housing. Population was at its housing for about 12 of the 33 minutes (mostly 3:11 to 7:56, 9:15 to 12:02 and from 30:45), with 8 to 23 idle and no open jobs. Three hovels stood by 8:52; the person had 24 (12 here).
• Food without a chain. Four orchards and one dairy farm, no wheat farm, mill or bakery, and it never read those entries. Orchard searches gave one site each at 6:44, 9:53 and 20:27 and none at 12:16, 13:33, 21:31, 21:56 or 22:19. It bought food seven times (below).
• Popularity. From 12:40 average taxes (−4) against food +1 and religion +2 let it drift from 80 to 55 by 31:00. The last two minutes cost 5 more points (0.5 of score) through mean taxes at 32:06 and a lapse in food variety; the model put the gain from mean taxes at about 40 gold.

What worked

• Iron and stone. The 143 iron and 152 stone in store are 4,353 of the 7,435; 60 more iron and 65 stone were sold.
• Jobs against the idle cap. Idle was 24 at 22:43 and again at 28:18, each time followed by woodcutters; population rose from 82 to 106 by 31:02.
• Selling to pay. 2,085 gold from stone, iron and wood left net worth unchanged and gave the cash for food at 22:42 and 27:30. The four chapels cost 1,000 gold; religion read +2 from 9:54 and +4 from 26:10, after the third (placed 21:31), and the fourth (26:09) changed nothing visible. High taxes at about 95 people paid 94 gold a month.
• Placement. All 80 reported placed stand in the final count; seven attempts were not placed, each with a status and remedy.
• Compaction and cache. One compaction at 19:56 (estimate 77,111 to 43,733 tokens, 22.9 s); the next call followed its handoff. All 69 gameplay requests read the cache, 97.0% of input tokens.

Market

195 planks (4 each, worth 1), 300 bread and 40 cheese (8 each, worth 4) were bought for 3,500 gold; 65 stone (455), 60 iron (1,380) and 250 wood (250) sold for 2,085. The purchases lost 1,945 of net worth against the sell price (planks 585, bread 1,200, cheese 160), 26% of the final value or 8.5 points of score; sales leave it unchanged. The 340 food units were bought at 9:25, 17:21, 20:00, 22:42, 27:07, 27:30 and 30:37.

Against the recording's gold and stock before and after each call, 16 of 17 replies agree (purchases within a unit, gold within what the town earned meanwhile). The exception is the 30:11 sale: the reply says 100 wood for 100 gold, but the recording shows 250 gone (576 to 334) and gold up 250, the known 1.7.3 undercount at 200 or more in stock; its stock and gold fields were right. The model saw it ("Wood dropped extra ~133") and let it go. The 17 calls took 387 game seconds; the 100 bread took 49.

Cost and speed

69 turns, 125 tool calls (59 find_sites, 17 build_structure, 17 market_trade, 13 building_info, 7 set_tax, 4 waits, 8 others), 3.54 million tokens and $1.271 at Kimi K3.1 Pro's list prices (preparation $0.007, play $0.821, compaction $0.034); the interview added $0.043. Play was 8,750 uncached input tokens ($0.018), 94,352 cache writes ($0.189), 3.30 million cache reads ($0.660) and 36,443 output tokens ($0.364). The run took 44.4 real minutes, 15.5 of them waiting for the model (13.5 s a reply, longest 34 s).

build_structure took 1,084 game seconds (17 calls, median 63 s) and trades 387, together 74% of the 2,000; waits took 128, taxes 54, expand_storage 42, set_workshop 25; reads took none. That is 13.6 s of game time per building placed by build_structure (80). This model's other games, same method: Oasis no arms 12.9 s (93 buildings), Desert 12.0 s (48) and 12.7 s (67), Sinkhole 15.7 s (106) and 14.9 s (129).

The interview

It put the gap down to productive capacity (the person's 202 people, 140% worker efficiency, bread chain and extraction) against its own 106 housing, heavy taxes, repeated food imports and "overinvested" woodcutters. Playing again it would farm and bake first, buy wheat, auto-sell bread, add a church and bad things, and avoid the 24-idle cap. It praised map_overview, find_sites and the readings, and asked for a bulk or auto-trade tool and a live net worth; no tool sets the marketplace's auto-trade, and no reading gave net worth.

These match the game: the totals, food imports (340 units, 2,720 gold, 1,360 of net worth), 37 woodcutters, 106 housing, the 24-idle cap and trades taking 387 s. These do not, or are not shown:

• "Absorbed workers who could have supported better chains": idle peasants were between 4 and 24 at every reading from 3:12 on (23 at the end) with open jobs at 0 (1 at two readings); the woodcutters were built to employ them.
• "Occupied useful land": at 22:19 its reasoning calls the five woodcutters of 19:00 a mistake on farmland, but the logs do not show a farm would have fitted there.
• "Taxed heavily enough to finish at 50 popularity": taxes were high or mean for 6.3 minutes (26:36 to 32:54) and average at the end; most of the fall (80 to 59 by 26:37) came under average taxes.
• "Verify production ratios early": it never read the entries for wheat farm, mill or bakery.

For the harness

• Sites that were refused. The hovel at (448,363) was the first site find_sites listed at 2:41 and the game refused it ("Too close to signpost to build"). At 4:40 the chapel (414,374) and hovel (417,382), also the first sites listed, were not placed ("no_free_spot_nearby"; the logs do not show why). Population sat at its housing for 4 min 45 s.
• Market undercount. The 30:11 sale above.
• Storage. Empty piles fell to 2 at 30:12 and 1 at 33:11; wait warnings cover popularity, food and gold only. It sold wood on its own reading.
• The 24 idle. At 19:56 its handoff said more buildings were not needed "merely to create jobs"; it found the cap itself at 23:36.
• Retry and watchdog. No retry fired, so this run does not show whether it works. The game's watchdog forced two collections after 20-second stalls (26:53 and 32:06). The game saved correctly.