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: Large Barren Desert 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 610d91be, series 20261017T171431-0e0530b1, 2026-10-17: Fable's no-arms desert game of the 1.7.3 batch, a fresh series with no memory of its others. Its weapons-counted game on this version (run 1adf1b13, series 20261017T164945-15a0128e) was played earlier that day; its desert games from earlier rounds are run edc045df (1.6.3) and run 86a73ff9 (1.5.6). This game counts on the website's leaderboard.

Result

A full-budget score: the run used all 2,000 game seconds. Net worth ended 145 above leaving the game alone, 3.0% of the person's 4,792, against −503 in the weapons-counted game. By the website's formula (80% net worth share, clipped to −1..2, 10% each population and popularity) the score is 11.6: +2.42 from net worth, +2.20 from 58 people and +7.00 from popularity 70; the weapons game's 1.3 is −8.40, +1.87 and +7.80. The stored goods are 71 iron (1,633), 194 stone (1,358), 20 ale (200), 20 pitch (200), and 4 bows and a spear, which count nothing. The town ended with 19 idle peasants, no free housing, no food and popularity falling (food −8, tax +1).

Against the whole starting package at the sell prices (6,560) the town ended 1,223 lower. By cause: stone +1,008 (144 mined), iron +1,978 (86 mined), taxes about +1,860 and weapons sold +2,925, against the workshops (−1,500), the 30 iron they used (−690), food bought and eaten (−4,540, with the start's meat) and planks (−2,265).

How the game went

Times are the game clock when a call began, read from the recording (the run clock quoted in observations reads 1 to 2 seconds later).

Mines and a weapons wing, 0 to 5 minutes. map_overview in its first reply (0:00) listed stone 83 to 106 tiles west of the keep, iron ore 57 to 69 tiles east, and no trees or farmland. Marketplace at 0:00, three hovels at 0:10, three iron mines at 0:23, three more stockpiles at 0:51 (the stockpile had no empty pile) and the fourth mine at 1:08: all four ore spots find_sites returned. An armory and fletcher's workshops by the stockpile found no room (1:22, 1:28); after find_sites at 1:50 it placed the armory, three fletcher's, a poleturner's and two hovels (done 2:39). Then 100 planks and 40 cheese (2:51), low taxes (3:25), a blacksmith's, an armorer's, two fletcher's and a poleturner's (3:38), 100 planks (4:17) and five fletcher's (4:41, done 5:02). Four more iron searches (3:27 to 8:53) found nothing.

Measuring, 5 to 19 minutes. 100 apples at 5:32. At 6:54 it had 46 people, 24 idle and no open job. find_sites at 6:55 gave three quarry spots near (412, 372); it bought 40 planks and placed one quarry and four ox tethers (7:05, done 7:28). Then waits of 60 to 120 s, weapon sales (8:30, 11:58, 14:27), planks, cheese and 80 bread. At 10:26 it wrote that a second quarry could help but it would hold off. The first stone reached the stockpile at 15:53, 8 min 48 s after the order. The context was compacted at 16:48. At 19:12, with 8 stone a minute in the last three minutes, it ordered two more quarries (19:38; one moved 6 tiles) and four tethers (19:57, done 20:14).

Selling and feeding, 20 to 33 minutes. It built nothing more; two hovels at 23:01 found no spot in view. Population reached 58 of 58 at 20:27. Its replies from 22:24 to 30:01 each sold weapons, bought food or planks and waited 100 to 150 s. From 24:48 the town ate cheese only (variety gone, tax −2 a month). A warning at 29:56 ("Popularity fell from 73 to 63") ended a wait after 108 of 150 s, and it set no taxes. The last armory sale was at 32:10 (500 gold); at 32:33 "Food is low: 17 in the granaries for 58 people" ended the next wait at once, and the last 45 s ran out with food at 0.

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 and pitch (400) are left out, so it reads 400 low until the End column, which is the game's own figure.

Weapons

It built them although they count nothing in store. Its opening plan says why: sell weapons "periodically so they turn into gold before the run ends". The armory and 14 workshops were all down by 5:02 and the rest of the game was selling. The armory never read above 34 of its 50; nine sale rounds (8:30 to 32:10, 200 to 500 gold each) brought 2,925, and 4 bows and a spear were left, since a sale takes lots of 5.

(Costs and inputs by prompt/game-mechanics.md: 2 wood a bow, 1 a spear, 1 iron a sword or suit; wood counted at 1, iron at 23.) The wing earned about what it consumed: +206, 3.4 score points. The planks were mostly bought: 505 of the roughly 750 that passed through the stockpile, at 4 gold. Charged with its share (529 of 750 planks) of the 1,515 lost buying planks, the wing comes to about −860 (an estimate). The metal workshops lost 130: 30 iron worth 690 became 900 of pieces for 300 gold and 40 wood. The last four sales (32:10 to 32:33) added 500, 8.3 score points.

What else the gold could do: the 1,500 bought 375 planks, nine quarry-and-tether sets at 40 planks each, if sites allowed (Opus and Sonnet each ordered 11 quarries in their no-arms games). Its three quarries and eight tethers, 100 planks, mined 144 stone worth 1,008. A quarry with four tethers also employs 7 people against 1 in a workshop, and 24 stood idle at every reading but one from 6:54 to 19:12.

The weapons-counted game built 17 workshops and 5 armories, made 210 pieces (2,840), sold 100 for 1,300 and held 110 worth 1,540. This game ended 648 above it: stone +1,358, gold +1,107, weapons −1,540, iron −437, other +160.

Beside its other games and the other models

Fable's four desert scores are 54.2 (1.5.6, no Armory offered), −0.7 (1.6.3), 1.3 (1.7.3, counted) and 11.6 (this game, no arms). Counted weapons were worth 2,737 and 1,540 in the two games that held them and nothing here; the score is 12.3 and 10.3 higher than in those two. That does not reach the 1.5.6 game, 42.6 points higher, which ordered 4 quarries at 2:02, ended with 12 quarries, 21 tethers and 586 stone (4,102), and had no Armory to build. Since the Armory was offered it built 13, 17 and 14 weapon workshops and 0, 0 and 3 quarries: the games differ most in quarries, not in the value of the weapons held.

Of the nine desert games finished in the batch, this is seventh by score and by net worth, and fourth of four models' no-arms games. The no-arms game beat the counted one for Opus (+48.0), Fable (+10.3) and Sonnet (+1.1) and trailed for Haiku (−3.3). Opus's lead is stone: quarries from 3:59, weapons only from 8:39, 760 stone (5,320, 60% of its net worth). Sonnet's no-arms game built no weapons and ordered quarries from 2:12. Sonnet's weapons-counted game built no workshops either: 4 mines, 5 hovels and 2,160 gold of food, and it ended at 6,448, 1,111 above this one; this game's purchases lost 3,675 of net worth, that one's 1,080.

What held it back

• Buying. 540 food cost 4,320 gold (86% of the starting 5,000), worth 2,160 on arrival and eaten by the end; 505 planks cost 2,020. The purchases lost 3,675 of net worth, more than the 2,925 the weapons brought. Its handoff put consumption at about 19 food a minute for 53 people.
• Jobs, then housing. 24 idle peasants held the town at 53 to 54 people from 8:05 to 19:56 with 5 places free; open jobs read 0 in every observation but two (0:45, 7:29). After the quarry sets idle fell to 18 or 19 and the 58 places filled by 20:27. For the last 13 minutes it placed nothing; the one hovel try (23:01, place_near) found no spot and find_sites was not tried.
• Stone late. One quarry at 7:05, two at 19:38; 144 stone in all, 10.7 to 16 a minute in the last four readings.
• Two wood a bow. Its plan ("1 wood → pike 18 / bow 15 / crossbow 30") and interview ("1 wood → 1 bow") take one wood a piece; the reference gives a bow or pike 2 and a crossbow 3. It called building_info for the mine and the poleturner's only. The stock fell 8.8 to 12.1 wood a minute in five stretches with no purchase or building, against 9.7 at two wood a bow and 5.7 at one (about 4 bows and 1.7 spears a minute); its handoff assumed about 5. Wood was 0 from 31:38.
• Popularity by low taxes and cheese. With low taxes (−2 a month, 3:25 to 29:56, about 1,860 gold) popularity fell from 100 to 62; the variety bonus was gone by 24:48.
• Waiting. 16 waits took 1,266 game seconds, 63% of the budget. From 7:28 to 19:38 it placed nothing.

What worked

• Iron first. Four mines on every site find_sites found, ordered by 1:08: 86 iron (1,978).
• Taxes. Low taxes for 26 minutes, with popularity never under 62.
• Tile targets. 8 build_structure calls placed 32 buildings; no call was refused and no request failed. All 40 placements reported as placed are in the final count.

Market

45 calls and 245 clicks. Bought: 310 cheese, 135 apples and 95 bread (4,320 gold), 505 planks (2,020). Sold: 105 bows, 45 spears, 15 swords and 15 metal armour (2,925). The purchases lost 3,675 of net worth against the sell price (food 8 for 4, planks 4 for 1); the sales added 2,925. These are units at the market's fixed prices. Gold: 5,000 − 4,320 − 2,020 − 1,500 + 2,925 + taxes = 1,946, so taxes were about 1,861; 43 paydays of 21 to 34 gold show in the recording outside the trade calls (1,342).

Cost and speed

30 turns, 99 tool calls (45 market_trade, 16 waits, 12 find_sites, 8 build_structure, 6 place_near, 3 observe, 9 others), 1.67 million tokens and $2.624 at the profile's prices ($10 per million input, $50 output, $0.25 cache read, $12.50 cache write): preparation $0.476, play $2.050, compaction $0.099. Play was output $0.833, cache writes $0.848 and cache reads $0.368. All 30 requests read the prompt cache, 95.6% of input tokens, with no provider routing. The agent run took 23.5 real minutes, 5.2 of them waiting for the model (longest reply 25 s); one compaction (19 s); 15 waits ran fast-forwarded (1,263 game seconds in 456 real).

Of the 2,000 game seconds: waits 1,266, the market 410 (1.7 s a click), building 259, other calls 29 (find_sites 13), five host top-ups of 5 s. build_structure took 189 s over 8 calls for 32 buildings, 5.9 s each; with place_near (53 s, 5 buildings) and expand_storage (18 s, 3), 40 buildings in 259 s, 6.5 s each. The same measure gives 6.5 s for its weapons game, 11.5 s for the 1.6.3 game and 8.0 s for the 1.5.6 game. The recording directory in events.jsonl shows the host on the game machine for all four.

The interview

Fable called its result barely above leaving the game alone (5,337 against 5,192), next to a person with four times its population: "a lot of activity that roughly broke even". It blamed a wrong value chain (fletchers at about 0.4 bows a minute, then no more expansion, against the person's wheat, flour and bread), stopping at 58 housing with 24 idle, quarries too late, food bought at half its value, and surveying and cautious waiting. Playing again it would commit in the first minute: hovels, quarries and tethers from minute 2, all the iron sites, mills and bakeries on bought wheat, inns and a church instead of food for popularity, and building to the end. It praised find_sites, tile build_structure, market_trade's net worth change and the per-building production, and asked for a note that food processors can run on bought inputs.

Against the game:

• Matches: 5,337 and 5,192; 250 against 58 people; about 0.4 bows a minute a workshop (0.3 to 0.4 in every window from 8:30); wheat at 23 (market_trade's price list: wheat 23/8, flour 32/10, bread 8/4); food at half its value (2,160 of 4,320); place_near hovels failing (1:36, 23:01); stone at 7.
• Does not match: "1 wood → 1 bow" is 2. "The first quarry around minute 16": ordered at 7:05, first stone at 15:53. "~8+ stone a minute per quarry": 8 was one quarry in one window (2.7 in the other); three quarries gave 0.9 to 5.3 each. "Stopped at 58 housing when 24 idle peasants halted arrivals": 24 idle at every reading from 8:30 to 19:12, with places free; housing was the limit from 20:27. Surveying: find_sites, building_info, inspect_building and observe took 22 game seconds (1%), waits 1,266. wait_and_observe "with until": never passed (0 of 16; screenshot: false on 13). Stale stats after trades: the host readings after trades (4:16, 19:56) agree with the trade results to within 4 gold once a payday or the workshops' gold is allowed for; its one "stale" remark (16:50) follows a wait. The compaction "made it more conservative": it ordered the quarries after it, and the handoff left them open.
• Not testable: the profit of mills and bakeries on bought wheat; the run never traded wheat or flour.

For the harness

• Weapon prices in a no-arms game. 1.7.3's rules leave out the prices, but market_trade's description lists them, and the opening plan quotes pike 18, bow 15, crossbow 30 before any market call.
• Inputs not shown. production lists what buildings delivered, not what they used; the fletcher's 2 wood a bow shows only in building_info and the reference.
• Warnings. The popularity warning came when the fall since the last observation reached 10 (73 to 63); readings of 83, 77 and 73 (22:23 to 27:35) gave none. The food warning came at 32:33, 47 s before the end.
• place_near came back partly or not placed in 4 of 6 calls (1:22, 1:28, 1:36, 23:01); the remedy names find_sites. The retry moved a quarry 6 tiles (19:38) and a fletcher's 1 (1:50).
• The only tool error was the last observe, answered "Session deadline" at 33:18.
• The game saved correctly at the end ("claude fable 5-1 1-7-3 610d91", 3,530,962 bytes, in this folder).