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, 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 the armory count toward net worth at the rules' prices, as since 1.6.0.
• Run 1adf1b13, series 20261017T164945-15a0128e, 2026-10-17: Fable's weapons-counted game of the 1.7.3 desert batch, a fresh series with no memory of any other game. Its no-arms game (run 610d91be, series 20261017T171431-0e0530b1) was played right after; its earlier desert games (run edc045df on 1.6.3, run 86a73ff9 on 1.5.6) were played on 2026-10-16.

Result

A full-budget score: the run used all 2,000 game seconds. Net worth ended 503 below leaving the game alone, −10.5% of the person's growth of 4,792. By the project website's formula (80% net worth, 10% each population and popularity, the net-worth share clipped to −1..2) the score is 1.3: −8.4 from net worth, +1.9 from population and +7.8 from popularity; this game counts on the website's leaderboard. It is 10.3 below its no-arms game, 2.0 above its 1.6.3 game and 52.9 below its 1.5.6 game. The 5,192 of leaving the game alone leaves out the part of the starting package that waits for storage room (about 1,150 at the sell price, per docs/benchmark.md), which this game took in by 2:31; against the whole package, about 6,500 by the rules, the town ended about 1,800 lower.

The stored goods are 90 iron (2,070), 88 bows (1,320), 22 spears (220) and 60 food (240): iron is 44% of net worth, weapons 33%. The town ended with 24 idle peasants, 24 free places in its hovels and popularity steady at 78 (food +2, tax −4, religion +2).

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).

A weapons wing, 0 to 5 minutes. map_overview in the first reply (0:00) listed stone 83 to 106 tiles west of the keep, iron ore 57 to 69 tiles east, and no trees, farmland or oil. The opening plan called wood turned into bows "the best margin on this map", at one wood a weapon, and quarries only "if stone proves practical". It placed the marketplace (0:00) and three hovels (0:09); an armory by the keep found no room in view (0:23). One build_structure call at 0:34 put up the armory, 4 fletcher's and 2 poleturner's workshops (37 s), and five more fletcher's and a hovel followed at 1:26. It bought 30 apples and 30 cheese (1:52, 2:01), sold the 20 pitch and 20 ale for 400 (2:31, 2:39) and bought 100 planks (2:47). Four more fletcher's (3:35) and a last fletcher's and poleturner's (4:35) made 14 and 3 by 4:48, 1,700 gold in all; gold was 2,443 at 5:03.

Iron, 5 to 9 minutes. At 5:03 it judged the workshops' margins thin, looked up the iron mine and the quarry in building_info, and placed four iron mines by tile on every site find_sites returned (5:22, done 5:48; searches at 5:48 and 8:55 around the other ore found none within 25 tiles). The first iron beyond the starting 15 arrived at 8:26. Armory piles (5:53), four hovels (6:06; a fifth was two wood short), 75 meat, apples and cheese (6:34 to 6:52), 100 planks at 8:20 after the workshops ran out of wood (a "Wood needed" notice, 44 s old at 8:02) and a chapel at 8:55 followed. Idle peasants read 24 at 8:02 and at every reading after; the population first reached 50 at 9:07.

Rations, taxes and waiting, 10 to 33 minutes. At 10:36 it set half rations and low taxes; popularity fell from 72 to 61, the warning ended the wait after 57 of 90 s, and full rations and no taxes followed (11:54, 11:57). With a religion factor of +2 (0 at 11:54) popularity was 78 at 13:30, and average taxes from then, 40 gold a month, held it there to the end. The only building after 9:06 was two armory piles (16:31). Its replies at 19:45, 23:01, 26:47, 27:33 and 30:54 bought planks or food, sold weapons, or only waited, and each ended in a wait of 150 to 200 s. The context was compacted at 26:46; the next reply doubted its trades (see Market), sold 20 bows and bought 40 planks and 90 food, and gold fell to 379 (a warning ended the next wait at 27:33). The last purchase was 15 planks at 30:54.

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. At 30:54 the armory held 76 bows and 16 spears (1,300), which puts the 30-minute figure near 4,100.

Weapons

It built an armory with the first workshops (0:34), four more armory piles by expand_storage (5:53 and 16:31; room for 250), 14 fletcher's and 3 poleturner's workshops by 4:48, and no blacksmith or armorer. It never called set_workshop, so the fletchers made bows and the poleturners spears by default, and it never called building_info for a workshop. The plan took a weapon to cost one wood; the table in prompt/game-mechanics.md says two for a bow and one for a spear, and the wood stock fits the table (46 to 10 between 27:33 and 30:54 while 17 bows and 6 spears were made, 40 by the table).

The wing made 148 bows and 62 spears: 88 and 22 held, 60 bows and 40 spears sold for 1,300. That is 2,840 at the sell price, from about 358 planks. The first pieces were in by 3:35. Once supplied, a fletcher made 0.3 to 0.4 bows a minute (the 14 together 4.3 to 5.7) and a poleturner 0.5 to 0.8 spears. The 1,540 held are 25.7 of the game's 1.3 points; with weapons worth nothing the same town would score −24.4.

From the table, the buildings cost 1,700 gold (17 workshops at 100) and 335 planks (workshops 20 and 10, armories 5). Against those and the 358 planks used, the pieces gained about 450 with planks at their sell price of 1. The planks were bought at 4 (590 for 2,360 gold), and the wing used about 693 of the roughly 830 the town used; priced as bought, that gain becomes a loss of about 900 to 1,300 (448 to 581 of the wing's planks were bought, depending on which of the start's 245 it used).

The same gold and planks could have gone to stone, which lay 83 to 106 tiles west and which the plan called too far; it ran no find_sites for a quarry. In its 1.5.6 game 12 quarries and 21 ox tethers cost 345 planks and gave 536 stone (3,752); in its no-arms game 3 quarries and 8 tethers gave 194 (1,358). The four iron mines, the only other producers, cost 80 planks and gave 75 iron (1,725).

Beside its other games and the other models

Weapons do not explain Fable's scores across versions: its 1.6.3 and 1.7.3 weapons-counted games, each a weapons wing and no quarry, scored −0.7 and 1.3, and the 1.5.6 game that scored 54.2 had no workshops and 12 quarries. What moved was where the starting gold went. The no-arms game still built 14 weapon buildings and sold 180 pieces for 2,925, added three quarries and ended 648 higher: 1,358 more stone, 1,107 more gold and 400 more in ale and pitch (this game sold them), against 1,540 less in weapons, 240 less in food and 437 less in iron. Of the 10.3 points between the two games, 10.8 are net worth and 0.3 population, less 0.8 for popularity.

In the batch so far (weapons counted / no arms) Haiku scored 18.9 / 15.6, Sonnet 27.6 / 28.7, Opus 19.1 / 67.1, Fable 1.3 / 11.6, and GLM 5.3 Flash −9.8 for its weapons game. This game is eighth of the nine played and fourth of the five with weapons counted. The other weapons games held 1,055 (Haiku), 1,095 (Opus) and 1,185 (GLM) in weapons; Sonnet built none. Opus has the nearest plan, 13 fletcher's and 3 poleturner's workshops and the same four mines, plus 7 quarries with 12 tethers; it ended 1,176 higher (5,865, 19.1) with 266 stone (1,862). In the 1.6.3 round this score would have stood eighth of nine, above Luna (−3.4) and below Kimi (2.7).

What held it back

• The opening wing and no quarry. 1,700 gold went into workshops in the first 4:48, 34% of the starting gold, and they took 17 of the town's 25 jobs, on a plan that took a weapon to cost one wood. The quarry was looked up once (5:03) and never searched for.
• 24 idle peasants from 8:02, and nothing built after 9:06 but armory piles. The rules say arrivals stop while 24 stand idle, and the town stopped at 50 people. At 8:02 and 8:55 its reasoning judged more workshops not worth their cost and idle peasants a drain; from 16:43 to the end, 16 min 37 s, it placed nothing.
• Food and planks bought. 535 food for 4,280 gold, 86% of the starting 5,000, counted 2,140 on arrival and 240 are left; 590 planks for 2,360 lost 1,770 at the sell price, most of them for the wing. It sold its 20 ale at 2:39 and looked at the inn only at 8:02, after the ale was gone.
• Taxes. They brought about 2,380 in all (the gold left after trades, sales and buildings). The half-ration test came with popularity falling from 72 to 61 in 78 s.

What worked

• Mines first, by tile. Four iron mines on every site found, placed from find_sites without a screenshot; they gave 75 iron, 2,070 at the end, about 0.7 to 1.0 a mine a minute (16 to 23 gold) against about 5 gold a minute from a fletcher.
• Placement and remedies. 28 of 29 build_structure placements went through; the one that did not, a hovel two wood short, named its fix. A fletcher blocked at its target was placed 4 tiles away and said so (1:26). No call was refused, and every placement reported as placed is in the final count.
• Warnings. The popularity warning (11:54) was answered in the same reply with full rations and no taxes; after the gold warning (27:33) it bought only 15 planks more.
• Food held to the end. The lowest reading after 3 minutes was 29 (23:15), and popularity ended at 78, where its 1.5.6 and 1.6.3 games ran out of food in their last minutes.

Market

33 market_trade calls (258 clicks, 369 game seconds, 18% of the budget, 1.4 s a click against 3.9 in its 1.6.3 game). Bought: 195 apples, 195 cheese, 145 meat and 590 planks for 6,640 gold; the purchases lost 3,910 of net worth at the sell price (planks 1,770, apples 780, cheese 780, meat 580; selling leaves it unchanged). Sold: 20 pitch and 20 ale (400), 50 stone (350), 60 bows (900) and 40 spears (400), 2,050 in all, which paid for the next purchases. Those from 26:55 on cost 525 of net worth.

At 26:47 the reasoning set the 599 gold of the 23:35 result against the 879 of the 26:46 stats and read the trades as not having gone through. They had: three minutes of taxes at 40 a month make the 280, and get_inventory at 26:47 matched the results (cheese 29, meat 28, apples 29, bows 75).

Cost and speed

28 turns, 88 tool calls (33 market_trade, 14 waits, 11 find_sites, 7 build_structure, 23 others), 1.60 million tokens and $2.384 at the profile's prices ($10 per million input, $50 output, $0.25 cache read, $12.50 cache write): preparation $0.450, play $1.839, compaction $0.095. The interview cost about $0.105 more, with 40,529 of its 43,458 tokens read from the cache. The run took 22.4 real minutes, 4.8 of them waiting for the model (longest reply 36 s); the context was compacted once (17 s). All 28 requests read the prompt cache, 96.3% of the input tokens.

Of the 2,000 game seconds the calls took 1,980: waits 1,312 (66%), market_trade 369, build_structure 174, expand_storage 47, place_near 32, the rest 46. build_structure took 174 s for 29 placements, 28 placed: 6.2 s a building. On the same measure its 1.6.3 game took 10.0 s (189 s for 19), its 1.5.6 game 7.8 s (365 s for 47) and its no-arms game 5.9 s (189 s for 32); all four ran with the host on the game machine (Fable's games, since 2026-10-16). The batch's other desert games, hosted the same way, ran 5.5 to 9.1 s a building by this measure, bar Sonnet's weapons game (22 s for 4).

The interview

Fable called its result poor: below leaving the game alone and under half the person's net worth, with 50 people against 250. It put the gap down to a low-yield chain (bows from bought wood, against the person's wheat, flour and bread), to food bought for a town where the person ran no rations and held popularity with inns, a church and a cathedral, to stopping at 8 hovels, to skipping the quarries at 100 tiles and to not considering fear buildings. Playing again it would measure a mill, bakery and fletcher first, set rations to none, grow population, build quarries with tethers and every iron mine early, and stop buying in the last minute. It praised tile placement, find_sites, expand_storage, market_trade and the stats, and asked for input-to-output ratios in building_info.

Claims checked against the run:

• Match: 0.3 to 0.4 bows a fletcher a minute and about 5 a minute for the wing (about 75 gold), food at 17 to 18 a minute (the compaction handoff), 8 hovels and 50 people, the stone at 83 to 106 tiles, an empty notebook and plan (run.json), one compaction. Those 75 gold covered about half of the 136 a minute that 17 food cost at 8, not "barely" all of it.
• until was not used. It called wait_and_observe with until handy; none of its 14 waits set it.
• The "stale" trade output was its own misreading at 26:47 (see Market), not output after the compaction.
• The conversion ratios. It says it never checked them. True for the fletcher: no building_info call for a workshop, and the plan assumed one wood a weapon.
• Not checked by the run: wheat at 23, bread at 4 and the mill's and bakery's ratios; it traded none.

For the harness

• The first armory by the keep found no room in view (0:23, no_free_spot_in_view), as in its 1.6.3 game; find_sites at 0:34 gave spots. In its no-arms game the armory and a fletcher's failed the same way at 1:22 and 1:28.
• Ratios live in building_info. The table gives bows 2 wood, crossbows 3, spears 1 and pikes 2, but no workshop was looked up, and the prompt's chain line has no ratios. The table has none for the mill, bakery or brewery, which the interview asked for.
• The game saved ("claude fable 5-1 1-7-3 1adf1b" in episode.json, 3,529,226 bytes, in this folder).