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 1.6.3 (commit 1d0c697), 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
• Run 9f070d6b, series 20261016T133433-0bba1ea0, 2026-10-16: Fable's game of the 1.6.3 round on this map, a fresh series with no memory of its other games. Its other Oasis game (run c7ad54c2, series 20261016T151706-833a56ad, version 1.5.1) was played later that day.

Result

A full-budget score: the run used all 2,000 game seconds. Net worth grew by 7,606 over leaving the game alone, 42% of the person's 18,293, against 7,230 (40%) in the 1.5.1 game. With 82 people and popularity 80 the score is 45.2 by the project website's formula (80% net worth, 10% each population and popularity, the net-worth share clipped to −1..2); the game counts on the website's leaderboard. The 1.5.1 game's 90 people and popularity 100 made 46.0.

The stored goods are 256 iron (5,888), 241 stone (1,687), 14 weapons worth 285 (9 bows and 5 crossbows), 101 wood (101) and 23 apples (92). Iron and stone are 86% of net worth. The town ended with 24 idle peasants, 8 empty places in its hovels and 23 food, which the game's own count calls one month's.

How the game went

Times are the game clock when a call began; a build or buy call keeps the game running for 20 to 80 seconds while it clicks, and the game is paused while the model replies.

The opening, 0 to 4 minutes. Its first calls were map_overview and find_sites for woodcutters. The overview listed two tree groves, 42 trees 62 tiles from the keep and 8 trees 27 tiles away, and its reasoning took trees to be scarce, so it built two woodcutters (on 5- and 12-tree spots) and a marketplace at 0:00, two hovels at 0:19, and bought planks: 60 at 0:40 and 85 at 1:59. Two iron mines, an apple orchard and a dairy farm came at 1:14, two more iron mines and two orchards at 2:51, two more hovels at 3:47. The first iron arrived at 3:50. At 4:13 the stockpile had no empty pile left, and it added three at 4:15. Gold fell from 1,000 to 420.

Mines and a weapons start, 4 to 14 minutes. Low taxes at 4:50, two more iron mines and two hovels at 4:55. At 6:31 it built two fletcher's workshops (100 gold each); a poleturner's workshop was refused for want of 9 gold, and gold touched 66 at 6:54. An armory beside the keep found no free spot (7:26) and went up at 7:53 with the first quarry and a second dairy farm; one fletcher was set to crossbows at 9:11. There were 24 idle peasants from 7:53, with 44 people in housing for 58. Two ox tethers and two iron mines at 9:45; 25 iron sold at 10:44 for 575 gold to buy planks; two more mines at 11:42 and two at 13:39 made twelve, with taxes at none from 13:31. At 14:37 find_sites found no iron mine site within 25 tiles of (418, 333), on the ore strip.

Stone and housing, 14 to 19 minutes. Two quarries and two tethers at 14:37, two orchards at 15:27 (a third found no room), a quarry, a tether and two hovels at 16:45 (housing 90 from 17:46), 20 bread bought at 17:53, and four more tethers at 18:30. The first stone arrived at 13:52; it stayed at 41 from 14:05 to 18:48 and then rose.

Taxes, then waiting, 19 to 33 minutes. It raised taxes to average at 19:13 (64 gold a month, popularity −4 a month), bought 45 apples at 19:41, and waited 90 s, 150 s, 200 s and 280 s. It set taxes back to none at 24:28, placed two ox tethers at 24:40, and built nothing after that: the last 8 min 40 s went on waiting, with 773 gold in hand and wood piling up to 101. Popularity fell from 96 at 19:13 to 54 at 24:32 and climbed back to 80; population stood at 80 from 19:11 and 82 from about 26 minutes. The host added a note after the waits of 4, 3 and 8 minutes in a row ("it may be time to do something else"); the model kept waiting. Its last observe, at 33:09, was answered "Session 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 285 in weapons is in the End value only).

Beside its other games

The two towns have the same shape: twelve or thirteen iron mines on the one ore strip, stone from the western deposit, five orchards, no bread chain, 24 idle peasants at the end. The 1.6.3 game ended 376 higher in net worth (gold +330, stone +504 and weapons +285, against iron −460, wood −235 and apples −48) and the 1.5.1 game ended 0.8 points higher in score: this game's net worth is worth 1.6 points more, its 82 people 0.4 less and its popularity 2.0 less.

The prompts differ in ways that show in the builds. The 1.5.1 rules said weapons and armour add nothing to net worth and gave no weights; the 1.6.3 rules list weapon prices, say net worth counts most, and say "New peasants stop arriving while 24 stand idle". So the 1.5.1 game built no weapon workshop and this one built two. The 1.5.1 game met the idle cap without being told: at 23:57 it noticed 82 people with 24 idle and housing for 122, placed eight woodcutters at 25:02 and 27:43, and population rose to 90 over the next five minutes. It also built a chapel, and spent a small bribe from 28:55 to 32:22 to carry popularity from 72 to 100. This game noted the idle peasants at 18:30 and afterwards placed only ox tethers.

It chose differently on wood: 2 woodcutters and 340 planks bought here, 14 woodcutters and 275 planks bought there. In the 1.5.1 game it ended with 336 wood, in this one 101.

Beside the other models. Among the other models' second games on this map, also 1.6.3, a score of 45.2 would stand third, behind Gemini 4 Argon (68.0) and GPT-6.1 Sol (54.6) and just above Claude Sonnet 5.5 (44.9), which had the higher net worth (10,107 against 8,826) but 26 people. By net worth alone it would be fourth, after Gemini (13,603), Sol (10,453) and Sonnet. Opus (28.9), Luna (25.9), Muse Wonder (23.7), Haiku (21.8), Kimi K3 (20.9) and GLM (15.3) are below. The best score of the earlier 1.5.x round on this map was 52.0 (Opus).

Weapons and armour. It built weapons but no armour. Two fletcher's workshops (200 gold and 40 planks between them) and an armory (5 planks) delivered 9 bows and 5 crossbows, 285 at the sell price, 3.2% of its net worth. No blacksmith's, armorer's or tanner's workshop was tried, and the one poleturner's workshop it asked for was refused for gold (6:31) and not tried again. Between the readings at 28:27 and 33:10 the twelve iron mines added 54 iron (1,242 at the sell price, about 100 a mine) and the two fletchers 3 weapons (60, 30 each). Its reasoning at 10:44 set workshops against mines and chose mines, at about 600 gold a mine for the rest of the game by its own estimate; at 16:14 it passed on a poleturner's workshop as not worth its cost.

What held it back

• No new production from 16:45. The iron strip was full at 14:37; after the quarry at 16:45 it added only ox tethers (six, the last two at 24:40). Its last two find_sites calls (19:26, 19:30) found no dairy farm or orchard site within 25 tiles of (475, 405) and (425, 400). From 19:41 it built two tethers and nothing else, and it ended with 773 gold and 101 wood unspent, after four long waits and three host notes about waiting.
• Raw materials only. Iron and stone are 86% of the net worth. Its gold came from taxes (low from 4:50 to 13:31, average from 19:13 to 24:28) and 25 iron sold; there was no bread, flour, ale or other processing. Its logged reasoning never mentions a bakery, mill, wheat farm, church or fear building, and it weighed workshops only against mines. The person named bread, sold through 57 bakeries, as one of its main sources of income.
• Idle hands and housing. At least 17 peasants were idle in every reading after 7:53 and 23 or 24 in each from 17:54. The rules of this version tell the model that new peasants stop arriving at 24 idle, and its reasoning at 18:30 noted "no new ones arriving"; it placed no hovel after 16:45 (housing 90) and the population stayed at 80 to 82. By the formula the 8 empty places were worth 0.4 points, so this cost little.
• Planks bought, not cut. 340 planks for 1,360 gold lost 1,020 of net worth against their sell price (12% of the final 8,826) and took 270 game seconds of market clicks (13.5% of the budget); they were about two thirds of the roughly 530 planks its buildings cost. Part of that bought the mines (at 10:44 it put each at about 600 gold), but 101 planks were left in store.
• Popularity ended at 80, not 100. The raise to average taxes at 19:13 cost 42 points (96 to 54 at 24:32) for about 720 gold; by the formula the gold is worth about 3.2 points and the 20 popularity points under 100 at most 2.0, so the raise was about even or a little ahead.
• A thin food supply. Apples were the only food from 21:55 ("1 of 1 foods eaten"), with 45 apples and 20 bread bought for 520 gold; food fell from 82 at 21:55 to 23 at the end.

What worked

• Twelve iron mines by 14:32. They dug 256 iron (5,888), 67% of the net worth; at 10:44 it chose mines over workshops and sold iron for planks to build more.
• Stone with tethers. Four quarries and eleven ox tethers gave 241 stone (1,687), against 169 from five quarries and four tethers in the 1.5.1 game.
• Every placement reported as placed is in the final count (12 iron mines, 11 ox tethers, 10 hovels, 5 orchards, 4 quarries, 2 dairy farms, 2 fletchers, 2 woodcutters, the armory and the marketplace; the stockpile counts 4 with the one at the start). Six placements did not go through: a dairy farm (6:31), a second dairy farm (7:53) and an orchard (15:27) for want of room, a hovel and the armory beside the keep for the same reason (the armory went up elsewhere at 7:53), and the poleturner's workshop for gold. No call was refused; the only tool error was the last observe, at the deadline.
• Storage held after 4:15: the full-storage warning appeared in one reading (4:13) and not again, and no model request failed.
• The prompt cache held on every gameplay request (below).

Market

340 planks (at 4, worth 1), 20 bread and 45 apples (at 8, worth 4 each) bought for 1,880 gold; 25 iron sold for 575. The purchases lost 1,280 net worth against the sell price (planks 1,020, bread 80, apples 180); selling leaves net worth unchanged. These are units at the market's fixed prices. Two plank purchases stopped short: 85 of 100 at 1:59 (no storage room, until the piles at 4:15) and 5 of 100 at 11:11 ("a purchase did not show in the reader within 5 seconds"). A lot of 5 took about 4 game seconds, so the 340 planks took 270 game seconds, and the nine market calls 362 in all. All calls other than waits took 1,183 of the 2,000 game seconds (build_structure 632 over 14 calls, market_trade 362, the 17 find_sites calls 11 in all), and the waits 754.

Cost and speed

28 turns, 69 tool calls (17 find_sites, 15 waits, 14 build_structure, 9 market_trade, 4 set_tax, 3 place_near, 2 observe, one each of map_overview, expand_storage, building_info, set_workshop and go_to_tile), 1.54 million tokens and $1.929 at the profile's prices ($10 per million input, $50 output, $0.25 cache read, $12.50 cache write): preparation $0.380, play $1.549, no compaction (the context ended at about 74,000 of its 120,000 budget). The interview cost about $0.963 more, with none of its 74,163 tokens read from the cache. The agent run took 30.7 real minutes, 4.6 of them waiting for the model (28 replies, about 10 seconds each). The 1.5.1 game cost $2.825 with 39 turns and one compaction, and its interview $0.128; the two games with their interviews cost about the same ($2.89 and $2.95).

Providers and the prompt cache

Every request went to Anthropic's API with no provider routing. All 28 gameplay requests read the prompt cache, 97.1% of their input tokens (1,455,994 of 1,499,219). The longest gap between a reply and the next request was 3 min 33 s, before the last one, so the cache never lapsed during play.

The interview did not read it. Its single request began at 14:11:58, 5 min 41 s after the last gameplay reply (14:06:17) and 5 min 52 s after that request began, past the 5-minute cache life, so all 73,205 tokens of context were written to the cache again ($0.915 of the $0.963). The run ended at 14:06:44 and the runner's log shows "Run completed" at 14:06:45 and the game saved at 14:11:58, 5 min 13 s later (all times UTC); the logs do not say why the save took that long. In the 1.5.1 game the gap was 4 min 30 s (its save took 3 min 26 s), and 59,889 of its 63,061 tokens were read from the cache.

The interview

Fable put its result at "roughly 6×" the do-nothing baseline and under half the person's net worth, and said the person found a value-multiplying chain (wheat, flour, bread, gold) and scaled it, with fear buildings, while it treated the game as "extract raw materials and hold them": 12 iron mines and a few quarries, "solid but linear". It added three faults of its own: waiting in the last 10 minutes, 24 idle peasants for most of the run, and housing never past 90. Playing again it would build the bread chain early and scale it (it "dismissed workshops as too slow but never evaluated bakeries"), keep peasants employed with hovels and jobs, use fear buildings with a church, build more tethers, quarries and woodcutters in place of buying about 300 planks, and keep building to the end. It found the tools good (find_sites with tile targets, build_structure retries, expand_storage, the stats JSON), but said each turn took 1.5 to 2 minutes of the budget, leaving about 20 decision points, and asked for per-building output. It called its main failure "strategic conservatism".

Against the game: the 24 idle peasants, housing of 90, 700+ gold and 100 wood unused, about 300 planks at a loss of 3 each (340 bought at 4 for a sell price of 1) and the missing bakeries all match, and "the last 10 minutes" is close (its last build was at 24:40, 8 min 40 s before the end, and from 19:41 it built only two tethers). Three statements do not. It built four quarries, not three. Its net worth is 7.2 times the baseline (the growth is 6.2 times). And it says exploratory or inspecting turns were punished heavily: its 28 replies do fall at 20 distinct game times, 1.7 minutes apart on average, but the replies use no game time (the game is paused for each) and inspecting calls cost almost none: the 17 find_sites calls took 11 game seconds in all and building_info and map_overview none. The budget went to building clicks (632 s), market clicks (362 s) and its own waits (754 s). Its wish for per-building output is not met in 1.6.3, which reports stored goods only; every observation since 1.7.0 reports production for each type of producing building (CHANGELOG, 1.7.0). It never traded wheat or built a bakery, so the wheat price it quotes (23) and "bakeries are cheap" are not checked by this run.

For the harness

• The map probe. Fable's 1.6.3 game ran 1.6.3's own probe, not the stale one built from 1.5.0's sources that the other models' 1.6.3 games on this map ran. It called map_overview once, at 0:00. The overview listed two groves, 42 trees at (477, 396) and 8 at (435, 385), and no cactus patch. The 1.5.1 game's overview (that version's probe) listed four, among them 23 at (349, 398) and 10 at (355, 365), the two patches the Opus 1.6.3 analysis names as cactus; the 1.5.1 game built no woodcutter at either. So the stale probe did not touch this game: what the overview showed it, few trees near the keep (8 at 27 tiles) and the large grove 62 tiles away, was real, and it bought planks on that reading.
• The interview missed the prompt cache (above): 5 min 41 s from the last reply to the interview, most of it the 5 min 13 s between the end of the run and the save. That made the interview cost about $0.8 more than the 1.5.1 game's.
• Market clicking takes game time. A lot of 5 units takes about 4 game seconds, so 340 planks used 13.5% of the budget; it is how the tool works, not particular to this run.
• Waiting notes did not stop the waiting. The host's note ("you have only waited and read for ...") came after 4, 3 and 8 minutes in a row and the model went on; food fell from 52 at 28:27 to 23 at the end, a month's by the game's count.
• The reader reported itself unavailable on the build at 13:39 (placement errors may be missing); all four placements were made. Fable's games were the first played with the host on the game machine itself, which changes nothing the game plays.
• The game saved correctly at the end ("claude fable 5-1 1-6-3 9f070d", in this folder).