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. Moonshot answered all 34 recorded requests (1 preparation, 32 gameplay, 1 compaction); none failed.
• Benchmark: Large Barren Desert construction (no arms), 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, with the game paused; what a run 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 store count nothing toward net worth; sold at a marketplace, they bring their gold. The host ran on the game machine.
• Run 207a50bd, series 20261023T180432-ebde4d9f, 2026-10-23: game 4 of Kimi K3.1 Pro's six ranked games, a fresh series with no memory of any other. Its weapons-counted game on this map is run a8a7d9e0 (series 20261023T172420-c995d1c5, −6.8), played 40 minutes earlier.

Result

A full-budget score: the run used all 2,000 game seconds. Net worth grew by 1,806 over leaving the game alone, 37.7% of the person's 4,792, against −972 in the weapons game. By the website's formula (80% net worth, 10% each population and popularity, shares from doing nothing to the person, the net-worth share clipped to −1..2) the score is 42.5: 30.2 points from net worth, 3.9 from 100 people, 8.4 from popularity 84. This game counts on the website's leaderboard. It is fifth of the eleven no-arms desert games by net worth and score, third in population (Gemini 4 Argon 136, GPT-6.1 Sol 105) and second in popularity (Sol 85). The gap to Sol, 98.5, is 56.1 points, 55.7 of them net worth.

The stored goods are 334 stone (2,338), 99 iron (2,277), 20 pitch (200), 18 ale (180), 25 food (100) and 3 wood: gold is 27% of net worth, stone 33%, iron 33%. The town ended with 100 people in 114 places, 24 idle and no open jobs, 25 food in the granaries (four kinds eaten) and 16 stockpile piles, 1 empty.

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). The game runs while a call clicks and is paused while the model replies.

The opening, 0:01 to 4:04. After map_overview (no trees, farmland or oil; stone 83 to 106 tiles from the keep, iron 57 to 69) its first reply noted "Interesting no trees!" and planned to spend the starting wood and buy the rest. One build_structure at 0:01 (90 s, 8 placements) put up the marketplace, three hovels and four iron mines. At 1:36 wood filled all four stockpile piles, so it added two stockpiles (1:37), then bought 50 meat (2:10) and 50 cheese (2:40) and placed three quarries and a tether (3:09, 51 s).

Taxes, services and jobs, 4:04 to 13:00. Taxes went from none to level 6 (4:04), with four hovels, an inn and two tethers (4:13, 96 s); popularity fell from 100 to 78 by 6:00. A chapel, two tethers and an armory (5:53), planks and food (7:13 to 8:24) and a granary followed. At 9:04 the town was 48 of 66 with 24 idle and no open job: "Pop48, idle24 cap -> growth stopped. I'll build jobs now." Three fletchers' workshops came at 9:05; the first quarry stone above the package's 50 arrived at 9:24 and the readout at 10:40 was 5.3 a minute.

A church and the dip, 13:08 to 20:41. Level 7 (13:08), two quarries and two tethers (13:34), the first crossbow sale (14:14). A refused wood purchase at 14:26 (see Market) left the next call (14:50) without wood for the inn. Popularity read 98 at 14:25, 89 at 16:00 and 71 at 19:07. The church, the inn and a sixth quarry came at 17:47 (66 s), five tethers at 19:31. Counting ale and pitch, net worth was lowest at about 3,890 at 18:32. The one compaction ran at 20:42 (39.8 s).

Quarries and tethers, 20:46 to 28:00. Level 8 (21:10), a quarry (21:19), five tethers (22:26), quarries at 24:32, 25:23 and 26:10 (the tenth). Popularity fell to 60 at 26:06 and it went back to level 7 ("Must set tax7 now -> upcoming0 (3+1+4-8=0), stable61"); a third inn and three tethers came at 27:55.

The last five minutes, 28:00 to 33:20. Crossbow sales, food and planks in small lots, two hovels (30:28; a tether was refused for wood, housing 114 from 30:57) and three tethers (32:31). At 32:59 it sold its last seven crossbows; a stone sale after them came back "Session deadline" (see Market). Population reached 100 at 33:10.

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; ale (10 each) and the 20 pitch (200) are left out, so it reads low by 400 to 540; the last column is the true net worth. With them added the town was under the 5,192 of leaving the game alone from about 9:20 to about 25:55 (an estimate; ale and pitch from the nearest observation).

Without weapons

It made weapons anyway, and sold them. Its preparation reply says it would "convert wood and iron into weapons for market sale rather than storing valueless armory stock", and its handoff says "armory weapons have no final value". At 0:01 it guessed "crossbow 1 wood ->30 good"; building_info at 4:04 gave 3 wood a crossbow and 2 a bow. It built five Fletchers' Workshops (100 gold and 20 wood each; 9:05, 11:21) and an armory (5 wood), set them to crossbows (two bows came out first, from the two placed at 11:21 before they were set at 13:08) and sold 35 crossbows and 2 bows in eight sales (14:14 to 32:59) for 1,080 gold, 1.0 to 2.0 weapons a minute in the readouts. The last 7 crossbows went at 32:59, so the armory was empty at the end.

The wing used 500 gold and 105 wood in buildings and about 109 wood of inputs (3 per crossbow, 2 per bow): 214 wood, which is 214 gold at net worth's 1 a plank or 856 at the 4 it paid. It ended between +366 and −276 gold (+6.1 and −4.6 points) ahead of not building it, and supplied five of the jobs the idle cap needed. At 14:26 it worked out that more workshops would take about 25 minutes to repay and built none; at 23:22 it dropped a late blacksmith ("value gain7 each only, ... cost200+wood, no."). Its weapons-counted game built 8 fletchers and 3 armories and kept 24 crossbows and 9 bows, 855 gold or 20% of its 4,220; counted at zero that town would score −21.1.

Beside its other games and the other models

Its three no-arms games scored 44.5, 43.3 and 42.5; the weapons games 37.3, −6.8 and 31.8. Per map: Oasis 40.9, desert 17.8, Sinkhole 37.6. Against its weapons game on this map (40 minutes earlier, same save and tools): 334 stone and 99 iron in store against 62 and 11 (that game sold 164 stone, 81 iron and the pitch for 3,211), 10 quarries and 28 tethers against 8 and 14, 13 hovels against 9, a chapel, a church and 3 inns against a chapel and 1 inn, and no wait_and_observe against 16. That game ended with taxes at level 10 (−20 popularity a month) and popularity 63 (−12); its lowest popularity was 39 (8:27), here 60 (26:06).

Quarries, tethers, stores and workshops are those standing or stored at the end, from each run's results.md; costs are as listed on the benchmark's table (this model's at Kimi K3.1 Pro's list prices). The four above it: Sol had 8 bakeries and a mill, Gemini 29 bakeries and 6 inns, DeepSeek 41 tethers and a blacksmith, Opus 35 tethers and 4 poleturners. This game had no processing chain and 28 tethers.

On the map score, the mean of a model's two desert games (from unrounded scores, here −6.84 and 42.45), it is seventh of eleven with 17.8: Sol 87.0, Gemini 70.3, Opus 43.1, DeepSeek 39.6, Sonnet 28.1, Muse Wonder 23.9, Kimi 17.8, Haiku 17.2, Luna 13.0, Fable 6.4, GLM −29.3. Over the three maps its mean is 32.1, level with Sonnet 5.5 and fourth of eleven.

What held it back

• Stone came late. The first quarries stood at 3:09 and the first stone above the package's 50 at 9:24; the readout was 5.3 a minute at 10:40 and at 14:25, 10.7 at 16:00, 13.3 at 20:41 and 21 to 27 from 24:48. Seven of the ten quarries and 23 of the 28 tethers were placed from 13:34 on. At the last readout's 2.3 stone a minute per quarry, ten more minutes of one quarry is 23 stone, 160 gold, 2.7 points. The stockpile had no room for stone from 32:47 (about 13 stone lost).
• Jobs. Idle peasants stood at 24 at 9:04 and were never below 18 afterwards, with no open job at any reading from 9:04 on except one at 16:00; at the end 24 of 100 people had no work. It saw the cap at 9:05 and each batch of jobs brought arrivals, but never enough open jobs to empty it.
• Food bought at a loss. 710 units for 5,680 gold, 2,840 of net worth lost against the sell price; about 740 left the granaries, 22 a minute. At 14:26 it worked out that "population at mean taxes may reduce net worth, but population score secondary". Rations stayed full all game (set_rations was never called) and it never called building_info on a mill, bakery or wheat farm.
• Services. The chapel, church and three inns cost 1,050 gold and 60 wood (17.5 points of net worth for the gold). They carried taxes at levels 6 to 8 (−6 to −12 popularity a month against +5 to +12 from food, religion and inns) and popularity ended at 84, worth 8.4 points.
• The refused wood purchases (Market) cost the inn three minutes (14:50 to 17:47) and a tether at 30:28.
• Cash. Gold stayed between 461 (22:17) and 1,170 from 18:00 to 29:30, so food went in small lots.

What worked

• Raw materials. 334 stone and 99 iron, 4,615 at store prices, 66% of the final net worth, from four iron mines in the first call and every quarry site find_sites offered (none more at 29:08). Nothing raw was sold before the last click.
• Taxes with popularity held. About 5,200 gold (5,160 by the recording's payday steps, 5,300 by difference), three-quarters of final net worth, at levels 6 to 8 over four changes; popularity never below 60, and back from 60 (26:06) to 84 with level 7, a third inn and religion +3.
• Jobs. Workshops, quarries and tethers were built to make jobs (9:05, 10:41, 13:08), and the population went 48 (9:04), 61 (15:00), 80 (24:00), 100.
• Compaction and placement. The one compaction (20:42) wrote a handoff that matches the 20:41 readout (gold 1,084, 72 of 98, 24 idle, popularity 80). 67 of 72 build_structure placements were placed (and 4 of 4 from expand_storage); the 5 refusals were two fletchers (touches_storage), an inn and a tether (wood) and a quarry (nothing fit).

Market

Purchases by the replies (units; gold): meat 170 (1,360), cheese 200 (1,600), apples 170 (1,360), bread 170 (1,360), ale 20 (400), planks 540 (2,160): 8,240 gold, 4,660 of net worth lost against the sell price. Sales: 35 crossbows (1,050) and 2 bows (30), then 50 stone in the refused last call (350). Gold reconciles: 5,000 start + taxes about 5,300 (by difference) + 1,430 sales − 8,280 purchases (the replies' 8,240 plus the two clicks below) − 1,550 for the workshops, chapel, church and inns = 1,900.

I checked all 48 market_trade calls against the recording (recording/inputs.jsonl clicks, frames.jsonl gold and stock). Each reply's clicks is the buy or sell clicks in the input log plus three more, and the stock and gold steps agree, with three exceptions. The wood purchases at 14:26 (asked 50) and 29:08 (asked 40) came back "not_traded", 0 units and 0 clicks, but one buy click landed in each: wood 34 to 39 at 14:34 and 18 to 23 at 29:16, gold −20, about a second after the click. And the last call (33:10) returned "Session deadline" with no figures, yet its click landed after the last frame: the final reading has 50 stone fewer (384 to 334) and 350 more gold (1,550 to 1,900). The model had asked for 20 (4 lots); the game sold 50 on the click. No other sale started above 9 in stock. The 48 calls and 409 clicks took about 890 game seconds, 45% of the budget.

Cost and speed

32 turns, 159 tool calls (48 market_trade, 40 find_sites, 18 build_structure, 18 observe, 16 building_info, 5 set_workshop, 4 set_tax, 3 expand_storage, 7 others), 1.61 million tokens and about $0.822 at Kimi K3.1 Pro's list prices (preparation $0.034, play $0.749, compaction $0.039); Moonshot reports no cost. The interview cost about $0.029 more (62,243 tokens, 59,483 cached). The agent run took 40.3 real minutes, 12.3 of them waiting for the model (23.0 s a gameplay reply, the longest 42.4 s). The 32 replies spent 18,865 reasoning tokens. All 32 gameplay requests read the prompt cache, 93.9% of input. No wait_and_observe ran.

Of the 2,000 game seconds, market_trade took about 890, build_structure about 851, expand_storage 67, set_workshop 52, inspect_building and observe 34 each. build_structure placed 67 buildings in 851 s: 12.7 game seconds a building (12.9 with the 4 from expand_storage). Its other games, build_structure alone: 13.6 on Oasis by the Sea (80 in 1,084 s), 12.9 on Oasis (no arms, 93 in 1,204), 12.0 on the desert with weapons (48 in 577), 15.7 on The Sinkhole (106 in 1,660), 14.9 on The Sinkhole (no arms, 129 in 1,920); all on the game machine.

The interview

It put 6,998 against 9,984 and 100 people against 250 down to direction: a quarry-and-iron economy with bought food as a running cost, against the person's bread from bought wheat, church, cathedral, inns and fear buildings, low taxes and no rations. Playing again it would make the wheat, flour and bread chain the main engine at once, build housing and jobs continuously, try zero rations and keep quarries as a side industry. It praised the read-only statistics, find_sites and the access checks; its obstacles were the game time that market and construction batches take and a purchase that "reported no reader change"; it asked for a net-worth and cash-flow reading and a production rate per building. Its biggest mistake, it said, was assuming bought food was necessary, and it had overvalued visible stockpiles.

• Match: 6,998 against 9,984; 100 and 250; goods 5,098 against 1,410; taxes at levels 6 to 8, full rations and a lack of jobs as the limits (idle 24 and no open job from 9:04; housing was the limit only around 29:08, 97 of 98); market and build batches taking 45% and 43% of the budget; stone and iron counting toward net worth (66% of it).
• "A purchase that reported no reader change and traded nothing": two replies said so; the recording shows one buy click landing in each (5 planks).
• Untested in the game: the wheat, flour and bread chain (no mill, bakery or wheat farm was ever priced; the plan follows the person's description) and zero rations (never set).
• "Rather than expanding tethers after their marginal returns had become uncertain": the readouts did not say that; stone went from 5.3 a minute (14:25, 8 tethers) to 21 to 27 (24:48 on, 20 to 28 tethers). Its own worry at 23:22, "Production quarry dropping despite tethers?", was a 3-minute window: 8.0, then 21.3.
• "My popularity was reasonable at the end": 84, from 60 at 26:06.

For the harness

• A refused purchase can have bought something. Both "not_traded" wood replies hid a landed click, and the same reply had already queued builds, so the inn (14:50) and a tether (30:28) were refused for wood. The model's remark at 16:01: "Uh market buy failed cost me time and resources missing".
• The last call. The stone sale returned "Session deadline" with no figures but its click landed (50 stone, 350 gold); the set_tax 8 and the 20 s wait_and_observe after it in the same reply never ran. The final reading is from the paused game, so the score is right.
• The full-pile warning came late. At 31:23 the status showed 1 empty pile and 94 stone room with 27 stone a minute arriving; "No empty pile" appeared at 32:58, after stone had no room from 32:47.
• Save. The end-of-run save needed a second attempt (attempts: 2) and was written ("kimi k3-1 preview 1-7-3 207a50", 3,548,063 bytes).