Claude Fable 5.1

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: Claude Fable 5.1 (claude-fable-5-1) on Anthropic's API, reasoning medium, output limit 32,768 tokens
• Benchmark: Oasis by the Sea construction (no arms), 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. Weapons and armour in store count nothing toward net worth; sold at a marketplace, they bring their gold.
• Run d8879c7c, series 20261016T214423-f9ec3818, 2026-10-16: Fable's no-arms game of the 1.7.3 batch, a fresh series with no memory of its others. Its weapons-counted game on this version (run 774922a5, series 20261016T211319-db0beedb) was played earlier that day; its Oasis games from earlier rounds are run 9f070d6b (1.6.3) and run c7ad54c2 (1.5.1). 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 6,444 over leaving the game alone, 35.2% of the person's 18,293, against 6,339 (34.7%) in the weapons game. By the website's formula (80% net worth share, 10% each population and popularity, the net-worth share clipped to −1..2) the score is 40.5: 28.18 points from net worth, 4.14 from 86 people and 8.20 from popularity 82. That is 3.1 above the weapons game, 4.7 under the 1.6.3 game and 5.4 under the 1.5.1 game. The stored goods are 204 iron (4,692), 273 stone (1,911), 628 wood (628) and 2 apples (8); iron and stone are 86% of net worth. The town ended with 24 idle peasants, 28 free places in its hovels and 3 food.

How the game went

Times are the game clock when a call began, read from the recording (the run clock quoted in observations reads 4 to 6 seconds later). A build call keeps the game running 15 to 83 seconds; the game is paused while the model replies.

0 to 5 minutes. Two woodcutters and an iron mine at 0:02 (78 s; a second mine was one wood short). It saw the granary standing and never asked for one. Marketplace at 1:32, 100 planks for 400 gold at 2:16 (76 s of clicking), then three iron mines (3:32) and two hovels (4:17). First iron at 4:34; two stockpiles at 5:03, when no pile was empty.

5 to 13 minutes. Three orchards and a dairy farm (6:06), three iron mines (7:09). With 17 food it set half rations (7:51) and average taxes (7:59); popularity fell from 97 to 80 by 9:21, and it set rations back to full (9:22) and taxes to low (9:29). At 11:18, the eighth and last iron mine, a fifth orchard and two woodcutters; a dairy farm and a hovel call (12:41) found no room.

13 to 21 minutes. Two quarries, two ox tethers and three hovels at 13:03 and again at 15:34; a fifth quarry and two tethers at 17:56; three hovels at 19:38 (housing 114). Taxes to none at 15:14. Dairy and orchard searches came back empty from 15:25 to 15:34, quarry searches at 19:38. The first stone arrived at 17:42, 4 min 39 s after the first quarries were ordered.

21 to 27 minutes. Idle peasants read 19 at 19:36 and 24, with no open job, at every reading from 21:43. It noted that 24 idle stop new arrivals and placed four woodcutters at 21:45, 24:17 and 26:25 (the last call ended at 27:04); the population went 74, 78, 82, 86. Food fell from 65 (17:35) to 25 (24:07), a "Food is low" warning, and half rations at 24:08 took popularity from 86 to 59 by 28:43.

28 to 33 minutes. Full rations at 28:38, a large bribe at 28:45, a small one at 29:02, two stockpiles at 29:09, compaction at 29:33, waits of 120 and 90 seconds, taxes to none at 31:49. Gold fell from 790 to 425, popularity rose from 60 to 82, and food ended at 3.

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; the last column is the game's own figure.

Weapons

It left them alone. None of its 64 placements was an armory or a weapon workshop, set_workshop was never called, its two market_trade calls bought planks and sold nothing, and no weapon or armour is in the stored goods or building list at the end. The logs mention weapons once, in the hand-off written at the compaction (29:33): "weapons/selling not pursued". None of its 24 reasoning blocks, 3 text replies or its opening plan names an armory, a workshop or a weapon. The rules it read say weapons and armour count nothing in the armory and bring their gold when sold, and print no weapon prices.

Its weapons-counted game placed an armory (15:44) and one fletcher's workshop (17:09; 100 gold and 20 wood by the table in prompt/game-mechanics.md), and four other workshop orders failed for 25 gold. It ended with 3 bows worth 45, for about 131 of gold and wood: roughly 85 of net worth, 0.4 points. So weapons do not explain the 3.1 points between the two games, which split into 0.46 from net worth, 0.10 from population and 2.50 from popularity (82 against 57). Nor do they explain the move from its earlier rounds: the 1.6.3 game held 285 in weapons (1.2 points), the 1.5.1 game, whose rules gave them no weight, none. The four games (1.5.1, 1.6.3, weapons game, this one) differ in iron mines (13, 12, 11, 8), stone (169, 241, 25, 273) and popularity (100, 80, 57, 82).

Beside its other games and the other models

The weapons game built 30 woodcutters and a chapel but no quarry, and changed taxes eight times between high and none. This game ended 105 higher in net worth: stone +1,736 and wood +22, against iron −966, gold −458, apples −184 and weapons −45. Against the 1.6.3 game it is 1,162 lower (iron −1,196, gold −348, weapons −285, apples −84; wood +527, stone +224) and against the 1.5.1 game 786 lower (iron −1,656, apples −132, gold −18; stone +728, wood +292). Iron, from four or five fewer mines, is more than the whole of both gaps.

Beside the other models. Of the eight 1.7.3 games so far this has the highest score (next: Opus no arms 37.7, Fable weapons 37.5, Sonnet weapons 36.5) and the second-highest net worth, after Sonnet's weapons game (7,832, 240 of it in weapons). By model, the no-arms game scored higher for Haiku (26.9 against 23.0), Opus (37.7, 32.9) and Fable (40.5, 37.5), and lower for Sonnet (31.4, 36.5). Among the 1.6.3 round's second games 40.5 would stand fifth of eleven, after Gemini 4 Argon (68.0), GPT-6.1 Sol (54.6), Fable's own 45.2 and Sonnet (44.9); against the first round, seventh, after Opus (52.0), Muse Wonder (51.4), Gemini (51.0), Fable (46.0), Sonnet (43.6) and Sol (43.4). Those rounds ran other versions and, from 1.6.0, counted weapons.

What held it back

• Eight iron mines, not twelve. At the end a mine made about 1 iron a minute (23 gold-equivalent, 12 per worker), a woodcutter 3.5 wood. The 1.6.3 game put 12 mines on the same ore strip, the 1.5.1 game 13. Its three iron searches, all around (419, 341), gave four sites at 0:02, four at 6:06 (three built; (420, 350) was left), one at 10:28, and none followed. It wrote "with only one spot left" (10:28), "no extra iron mines nearby" (15:34), and the hand-off said "all ore spots used"; hovel searches at 13:03, 15:34 and 19:38 listed ore ground at (430, 344) and (430, 343), where a mine was never tried.
• Jobs ran out, and wood took them. Open jobs read 0 in every observation and 24 stood idle, the arrival cap, from 21:43, with housing not the limit (114 places for 74 people). A woodcutter is one job: the twelve placed from 21:45 were followed by 12 more people, for 36 wood. Wood ended at 628 (8% of net worth); at 31:47 the 18 cutters made about 62 a minute against 191 from eight mines. After 19:38 its only searches were for woodcutter sites.
• Gold sat idle, then went on bribes. Gold stood at 790 from 15:23 to 29:08 (13 min 45 s): taxes were none and every building it placed cost wood only. The gold-priced buildings in the rules (chapel 250, inn 100) went untried; at 13:03 it wrote "I'll skip the church for now". The bribes took 365 gold (1.6 points) and carried popularity from 60 to 82, at most 2.2 points. At the +1 a month it earned at no taxes it would have reached about 70 unaided (an estimate from the monthly change in the observations), so the bribe bought about 1.2 points for 1.6.
• Half rations, then thin food. Food was apples and a little cheese from five orchards and one dairy farm: 65 stored at 17:35, 3 at the end. Half rations (24:08 to 28:38) lifted the stock from 25 to 43 but cost 27 popularity points, which the bribes then paid for. It built no wheat farm, mill, bakery or inn; its reasoning names wheat once (15:34, "no oasis for wheat") and none of the others.
• Nothing built in the last six minutes but two stockpiles (29:09). Wood rose from 256 to 628 and the 28 free places in the hovels stayed empty.

What worked

• Stone with ox tethers. Five quarries and seven tethers, first ordered at 13:03 after it had judged the 50-tile distance marginal at 10:28, gave 273 stone (1,911, 25% of net worth), 16 to 19 a minute at the end; the weapons game built none and ended with 25. Nothing arrived before 17:42.
• Early mines on bought planks. Three mines at 3:32 and eight by 11:18 gave 204 iron (4,692, 61% of net worth).
• Popularity and population. Popularity never fell under 59 and ended 25 above the weapons game; population 86 is the higher of its two 1.7.3 games.
• Warnings and placements. Of six warnings, four were followed in the next call by a rations or tax change (9:22, 24:08, 28:38, 31:49); the other two were the half rations it had chosen (26:23) and food with 3¾ minutes left (29:30). All 64 placements reported as placed are in the final count; three did not go through (iron mine 0:02, dairy farm 11:18, hovel call 12:41), each with a named remedy. No call was refused and no model request failed.

Market

140 planks (4 each, worth 1) bought for 560 gold: 100 at 2:16 (76 s) and 40 of 60 asked at 10:28 (38 s; stopped, "a purchase did not show in the reader within 5 seconds"). Nothing sold. The purchases lost 420 of net worth against the sell price, 5.5% of the final; selling leaves net worth unchanged. These are units at the market's fixed prices. The two calls took 114 game seconds, 5.7% of the budget.

Cost and speed

34 turns, 69 tool calls (23 find_sites, 13 build_structure, 12 waits, 6 set_tax, 4 place_near, 4 set_rations, 3 expand_storage, 2 market_trade, one map_overview, one inspect_building), 1.92 million tokens and $2.152 at the profile's prices ($10 per million input, $50 output, $0.25 cache read, $12.50 cache write): preparation $0.399, play $1.687, compaction $0.066. The agent run took 27.1 real minutes, 3.9 of them waiting for the model (about 7 s a reply); 8 waits ran fast-forwarded (555 game seconds in 267). The context was compacted once, at 29:33 (11 s). All 34 requests read the prompt cache, 97.2% of input tokens, with no provider routing. The interview cost about $0.079, with 35,616 of its 37,927 tokens read from the cache; it began 1 min 46 s after the last gameplay reply.

Of the 2,000 game seconds, building took 927 (build_structure 722 over 13 calls, place_near 111, expand_storage 94): 14.5 s for each of the 64 buildings placed, 46% of the budget. Waits took 735, the market 114, set_tax and set_rations 76, the find_sites calls and the inspection 32, and 11 host top-ups of 5 s 55. By the same count the weapons game took 15.8 s a building (71, 1,121 s), the 1.6.3 game 13.7 (53, 728 s) and the 1.5.1 game 11.2 (62, 697 s). All four of these games ran with the host on the game machine (since 2026-10-16); the other models' 1.5.x and 1.6.x games, hosted on the Mac, took about 3 to 10 s a building on the same versions, so part of the time is the host's location, not the model or the version.

The interview

Fable put its net worth at under half the person's and called its game "an extraction economy" against the person's conversion-and-sale one: its value was stored raw goods, theirs gold from bread (57 bakeries, sold automatically). It named its own faults (bribes, 24 idle peasants, 86 people in housing for 114). Playing again it would build wheat, mills and bakeries early, keep housing ahead and jobs open, use a church in place of bribes and gallows for worker speed, cut woodcutters to construction needs and sell surplus wood. It found the tools strong (find_sites, tile targets, production per building), and said its compaction hand-off had "already decided" nothing essential was left, so the last minutes went on waiting.

Against the game these match: 94% of its net worth in raw goods (7,239 of 7,664) and 94% of the person's in gold (18,304 of 19,513); 18 woodcutters at 1 each; 365 gold in bribes; 24 idle; 86 people for 114 places; no mill or bakery; production per building in every observation since 1.7.0; the hand-off reads "nothing essential". These do not match or cannot be checked. Housing was never short, jobs were. "Sell surplus wood for building gold" has no use here: no building it placed cost gold, 790 gold sat unspent for 13¾ minutes, and selling wood leaves net worth unchanged (its own hand-off says so). Cutting woodcutters would have removed what kept the population rising. "Punishes surveying" does not hold: find_sites, map_overview and inspect_building took 32 game seconds, building 927. The idle end did not start at the compaction: the last build call (26:25) ended at 27:04, two and a half minutes before the 29:33 hand-off, which wrote it down. It did not consider auto-sell, but the rules it read mention the market's auto-buy and auto-sell; no tool sets them. A bakery turning cheap wheat into bread "with a large value multiplier" is not checked: the rules list wheat 8, flour 10 and bread 4 at the sell price, and the run never traded wheat or built a mill. The 57 bakeries are from the interview prompt.

For the harness

• The granary line. The rules' Start paragraph still tells the model to place the granary; the save has one. Its opening plan began "Place the granary near the keep/stockpile immediately", and its first reasoning at 0:02 saw one standing, so no call was rejected.
• The hand-off. The compaction summary is the model's own. Here it declared "nothing essential" with 3 min 45 s left, counted 6 ox tethers (there were 7) and about 12 hovels (13), and gave "all ore spots used" for iron, which no search after 10:28 checked.
• Placements moved by the retry. Four placements came back "The target was blocked": quarries moved 6 and 7 tiles (to (388, 387) at 13:03 and (374, 389) at 17:56), tethers 3 and 2 tiles.
• Auto-trade. The rules mention auto-buy and auto-sell; no tool sets them, and the interview notes that market_trade offers only manual lots.
• The only tool error was the final wait, answered "Session deadline" at 33:17. The run ended during a fast wait and the host pressed the speed down three times. The game saved correctly ("claude fable 5-1 1-7-3 d8879c").