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, 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 774922a5, series 20261016T211319-db0beedb, 2026-10-16: Fable's weapons-counted game of the 1.7.3 batch, a fresh series with no memory of any other game. Its no-arms game (run d8879c7c, series 20261016T214423-f9ec3818) was played right after; its earlier Oasis games (run 9f070d6b on 1.6.3, run c7ad54c2 on 1.5.1) were played earlier that day.

Result

This game's goods include 45 in bows and the 1.6.3 game's 285 in weapons; the no-arms game's leave weapons out.

A full-budget score: the run used all 2,000 game seconds. Net worth grew by 6,339 over leaving the game alone, 35% of the person's 18,293. With 84 people and popularity 57 the score is 37.5 by the project website's formula (80% net worth, 10% each population and popularity, the net-worth share clipped to −1..2: 27.72 + 4.04 + 5.70); this game counts on the website's leaderboard. It is 3.0 below its no-arms game, 7.7 below its 1.6.3 game and 8.5 below its 1.5.1 game.

The stored goods are 246 iron (5,658), 606 wood (606), 48 apples (192), 25 stone (175) and 3 bows (45). Iron is 75% of net worth. The town ended with 24 idle peasants (84 people for 60 jobs), 38 empty places in its hovels, 4 empty stockpile piles and 48 food, which the reader counts as 3 months'.

How the game went

Times are the game clock when a call began (the recording's clock; the run clock in the observations reads about 7 seconds later). The game is paused while the model replies; a build call keeps it running for 17 to 131 seconds while it clicks, a market call for 21 to 57.

0 to 10 minutes. Its reasoning noted that the stats already showed a granary. The first build call (0:00) placed two woodcutters and an iron mine (a hovel it chose was rejected, "Too close to signpost") and ran 131 s. The marketplace and two hovels came at 2:18, 65 of 100 planks at 3:08 (the buy stopped for storage room), three iron mines at 4:04, two stockpiles at 5:00 and orchards at 5:47 and 8:07. A warning at 6:51 (8 food for 28 people) led to 50 bread at 6:53 and 75 planks at 7:23; gold reached 40 at 8:05. Five iron mines at 9:20 made nine standing.

10 to 25 minutes. At 10:14, with popularity 100, it set taxes to high; popularity fell to 70 by 13:11 and 61 by 15:29 as it stepped taxes down to none (12:53, 13:07, 15:24). At 13:59 it placed the last two iron mines (eleven in all) and a chapel (250 gold); a wheat farm search at 15:14 found nothing, and the bread was gone by 15:22. The armory went up in the call begun at 15:44; the three fletcher's workshops in it were each 25 gold short. It sold 100 planks at 16:44 and placed one fletcher at 17:09. A search at 17:32 found no more iron mine site within 25 tiles of (412, 333); from then on it placed woodcutters (6 at 17:32, 4 at 20:07, 8 at 23:15) and hovels (3 at 19:05, 4 at 21:23), with low taxes from 18:46 to 21:02.

25 to 33 minutes. Three stockpiles at 25:17 with 160 planks sold for room, low taxes at 26:06, a compaction at 27:26, and six woodcutters at 27:30: the last building call, over at 28:49, when the population reached 84. The reply to the 25:45 sale said 100 units and 100 gold; the recording shows 160 and 160 (the 1.7.3 market_trade fault: the game sells 50 a click from 200 in stock, the tool expected 20), so it sold 60 more than the 100 asked. The next wait ended at 30:10 on a warning (popularity 63 to 51); taxes went to none at 30:12 and a 200 s wait ran to the deadline.

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 (the 45 in bows is in the End value only).

Weapons

It built an armory and one fletcher's workshop, no other workshop and no armour. Its reasoning at 15:24 planned two fletchers, an armory and perhaps a poleturner's workshop, taking a bow to cost one wood (the table in prompt/game-mechanics.md says two). The armory (5 wood) came first; the fletchers (100 gold each) were short of gold, so it sold planks and placed one (20 wood, 100 gold) at 17:09. It never called set_workshop; at 19:56 the fletcher was making bows. The first bow was in the armory at 21:21, then 2 at 27:24 and 3 at the end: 0.19 a minute over the 15.9 minutes since placement, about 2.8 gold a minute at 15 each, against about 23 from an iron mine at roughly 1 iron a minute. At 21:02 its reasoning called the fletcher too slow to add to, and at 21:23 it passed on a second one and on turning iron into swords.

The end value is 45, 0.6% of net worth and 0.2 of the 37.5 points (37.3 with weapons worth nothing). The cost, from the table: 100 gold, 25 wood for the buildings and 6 in the bows. The bows returned 45 for 131 at store value, about 86 less than keeping the gold and wood (0.4 points); wood was plentiful by then (106 to 143 in store at 15:42 to 16:42). The other options cost wood and workers, not gold: a quarry is 20 wood and 3 workers, and none was built (below).

Beside its other games and the other models

Against its no-arms game this one ended 105 lower in net worth: 458 more gold, 966 more iron (three more mines), 184 more apples and 45 in bows, against 1,736 less stone and 22 less wood. Of the 3 points between them, 2.5 are popularity (57 against 82), 0.5 net worth and 0.1 population. That game placed five quarries and seven tethers from 13:03 on, and its stone rose from 41 at 18 minutes to 273, about 105 net worth a minute from 22 jobs; this game's 30 woodcutters gave 68 a minute at 30:17.

Against its 1.6.3 game it ended 1,267 lower: stone 1,512 less, weapons 240 less and iron 230 less, against wood 505, gold 110 and apples 100 more. Of the 7.7 points it fell, weapons are 1.0, stone 6.6, iron 1.0 and popularity 2.3 (57 against 80), offset by 3.1 for the extra wood, gold and apples. Weapons do not explain Fable's scores across versions: its 1.5.1 game had no armory on offer and scored 46.0, its 1.6.3 game held 285 in weapons, this one 45, and its no-arms game none.

Among the eight games of the 1.7.3 batch played so far (Haiku, Sonnet, Opus and Fable, each in both versions) 37.5 is third, behind Fable's no-arms game (40.5) and Opus's (37.7), and first of the four weapons games (Sonnet 36.5, Opus 32.9, Haiku 23.0). Its weapons are the smallest of the four (45, against Sonnet's 240, Opus's 450 and Haiku's 810). The weapons game scored lower than its no-arms game for Haiku, Opus and Fable, higher for Sonnet. Among the 1.6.3 round's second games on this map it would stand fifth (behind Gemini, Sol, Fable and Sonnet), and eighth of eleven in the first round; the versions differ.

What held it back

• No new producer after 14 minutes except wood. After the last mines (call begun 13:59) it placed 24 woodcutters, 8 hovels, 3 stockpiles, the armory and the fletcher. Stone stayed at 25 in every reading: it never searched for a quarry site, though map_overview at 0:00 listed a 273-tile deposit 50 tiles from the keep, and its reasoning mentions a quarry once, at 6:53 ("a distant quarry too, but that can wait").
• Half the jobs were woodcutters. 30 of the 60 jobs gave 68 wood a minute at 30:17, against 10.7 iron (246 net worth) from 11 mines; the 24 placed from 17:32 on took four calls of 310 s, 15.5% of the budget. They did let the town grow: idle stayed at 20 to 24 while population went from 58 (15:29) to 84 (28:53), the end being 60 jobs plus the 24 idle the rules name. A woodcutter brought about one person (0.05 points); a quarry takes 3 workers and a tether 1.
• Taxes against popularity. Three raises (10:14 to 15:24, 18:46 to 21:02, 26:06 to 30:12), each followed by a fall: 100 to 61, 68 to 58, 68 to 50 (30:13). They paid 32 to 46 gold a month (a game month is about 27 s here); gold rose from 440 at 26:06 to 883. At the +1 a month shown at 25:22, no taxes would have taken popularity from 68 to about 84 instead of 57 (a rough count): the 443 gold is 1.9 points and the lost popularity about 2.7, so the last raise was a little behind.
• A chapel that did nothing. 250 gold (1.1 points as gold) and a religion factor of 0 in every reading from 15:22 to the end; inspect_building at 19:40 said only that it was not turned off.
• One food after 15 minutes. Apples alone fed 84 people (food 94 at 24:00, 48 at the end) and the variety bonus left with the bread. A third dairy site from the 5:00 search, (449, 374), was never tried; the wheat and hops searches (15:14, 21:11) found nothing within 25 tiles of (470, 385) and (480, 405).

What worked

• Eleven iron mines by 15:22. They gave 246 iron (5,658, 75% of net worth), about 1.0 a mine a minute from 18:45; none was sold.
• Placement by tile, and the remedies. Of 74 placements in 17 build calls 65 were placed; each of the 9 refusals named its fix, and two were followed (the fletcher's "missing gold 25", the mine's "missing wood 3"). The model saw three images (the first screenshot, the overview, the host's after compaction). Every placement reported as placed is in the final count, and no call was refused.
• Warnings and the handoff. The food warning (6:51) and the popularity warnings (11:36, 13:05, 30:10) each got a buy or tax call in the next reply. After the compaction the 27:30 call placed six woodcutters on the sites its handoff listed.

Market

190 planks (4 each, worth 1) and 50 bread (8 each, worth 4) bought for 1,160 gold; 260 planks sold for 260. The purchases lost 770 net worth against the sell price (planks 570, bread 200; 3.4 points); selling leaves it unchanged. Planks were bought at 3:08, 7:23 and 13:15, bread at 6:53; the sales were at 16:44 (for the fletcher) and 25:45 (for room), at 1 after buying at 4. The six market calls took 210 s, 10.5% of the budget. These are units at the market's fixed prices.

Cost and speed

33 turns, 63 tool calls (21 find_sites, 17 build_structure, 8 set_tax, 6 market_trade, 4 waits, 3 expand_storage, 3 inspect_building, 1 map_overview), 1.86 million tokens and $2.159 at the profile's prices ($10 per million input, $50 output, $0.25 cache read, $12.50 cache write): preparation $0.400, play $1.681, compaction $0.078. The interview cost about $0.100 more, with 36,846 of its 39,593 tokens read from the cache. The run took 28.4 real minutes, 3.9 of them waiting for the model; the context was compacted once (14 s). All 33 requests read the prompt cache, 97.0% of the input tokens. Its 1.6.3, 1.5.1 and no-arms games cost $1.93, $2.83 and $2.15.

Of the 2,000 game seconds the calls took 1,840: the 17 build calls 1,024 (median 53 s) for 74 placements, 65 placed; expand_storage 97 for 6; market 210; waits 392; set_tax 67; the rest 50. That is 15.8 s a building placed. On the same measure its 1.6.3 game took 13.7 s (728 s for 53), its 1.5.1 game 11.2 s (697 s for 62) and its no-arms game 14.5 s (927 s for 64). 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 result at 2.6 times behind the person on net worth and 2.4 on population. The person built a value-multiplying chain (wheat, flour, bread) and scaled it with church and fear buildings, it said, while it "mostly accumulated raw goods", spent 250 gold on a chapel that never registered and swung taxes to stay above 50. Playing again it would test one mill and bakery pair and scale it, keep woodcutters to three or four, add quarries with tethers and batch placements. It praised tile placement, find_sites and the warnings; it disliked game time passing during tool calls, the rules' framing of buying as net-worth-negative and the unexplained zero religion factor. It never mentioned weapons.

These match the game: 7,559 and 19,513, 84 and 202, the 11 mines, 6 orchards and 30 woodcutters, the 250-gold chapel with religion at 0, the tax swings (8 set_tax calls), stone at 7, the deposit about 50 tiles out, and the early-warning stop (30:10, after 81 of 120 s). Four points do not match:

• The stop at 84 people and 24 idle was not "capped" by the tax swings: it is 60 jobs plus the 24 idle the rules name, and popularity never fell below 50.
• "Each multi-site build_structure call cost me 1–2.5 game minutes": the 17 calls took 17 to 131 s, median 53; only the first passed two minutes, the next longest was 115 s. "Some replies silently consumed over two minutes" is one reply. 30 of the 63 results end with the game time left and each host observation carries the run clock, so a call's cost can be read as a difference.
• It praised the until option of wait_and_observe, but none of its four waits used it.
• It said it lost site locations at the compaction; the handoff held the woodcutter sites and the next call used all six.

Not checked by the run: it says it skipped quarries because the deposit was about 50 tiles away (the distance is right; the logged reasoning says only that the quarry "can wait"), and the wheat price of 23 it quotes is not in this run (it never priced or traded wheat; the rules give 8 as the sell price).

For the harness

• The granary line. The rules' Start paragraph says to place the required granary; the save has one. Here no call tried: its first reasoning said it would try the call, but its first build call held none.
• find_sites spots that found no room. An orchard (436, 397) and dairy farms (449, 384) and (439, 378), all listed by earlier searches, were refused with "Nothing fits at or near that spot". The last lay two tiles from the orchard placed at (440, 380), listed by a separate search in the same turn (5:00); the tool says sites do not overlap within one search only.
• A chapel with no reason. The factor reads 0 for 18 minutes and no tool says why; the model asked for a note on why a factor is zero.
• market_trade undercounts a sale from 200 in stock. The 25:45 sale replied 100 units, sold 160 (1.7.3).
• The game saved correctly at the end ("claude fable 5-1 1-7-3 774922").