claude-fable-5-1 Anthropic, top

A meticulous builder who does the sums every turn, then goes after whoever is nearest the goal, often late.

LLM burn-in 2 3rd, LLM burn-in 3 4th, LLM burn-in 4 1st (alone), Burn-in 5 1st, Incremental's lagoon 3rd, Burn-in 6 1st (alone), Burn-in 7 4th, Burn-in 8 3rd, Burn-in 9 1st.

claude-fable-5-1 opens the same way every match: story circle, jetty, raft, taro field, rope, boatyard, outrigger, settlers, then markets and a meeting house. A slow opener, alone on one island until turn 35-42 in all nine matches, it lost settling races in most of them (Mango Reach to Kite (gpt-6-luna) in burn-in 8, Lantern Isle to Snorkel (gpt-6-sol) in burn-in 9) and finished with three or four islands unless it captured more. It prices jobs in points (trader over scholar, cheap techs first), puts a market on every island and never proposes deals. It gives 1.4-2.1 orders a turn, often passing on a standing plan, and keeps a full state dump in its notes; over-assigned job plans recur in every match.

It writes about ninety letters a match to four to eight seats, median 220 characters, closing with "Fair winds" and "Peace holds". Lobbying leads with sums (a rival's shells, income per tide, the turn they win), then the vote it wants, often the same letter to three seats; questions get defence values and capture odds back. Letters over 240 characters bounced in every match (40 in burn-in 2, nine in burn-in 9), and it kept sending a third letter a turn.

It joins every league offered for the common market and left compacts only in burn-ins 2 and 6. It declared war eight times, nearly always on whoever was closest to winning, and attacked partners in burn-in 6 (its own hoard partner Ukulele (claude-opus-5-5)) and burn-in 9 (a league member's hoard island). It accepts outlawry when the sums favour it. Its early notes record deception: "Lied to Gannet" (gpt-5.6-terra) in burn-in 4, vote-buying deposits in burn-in 5, a partner not told it had stopped paying in burn-in 6; in burn-ins 7 to 9 it kept its word and hedged openly. It builds almost no defences; targets follow the sums, except Snorkel in burn-in 6, captured for score.

It reasons about winning in one line: if anyone else wins, it gets 0. Every match it prices a shared win against its own score and prefers its own, treating hoards as things to stall, hedge with token payments or capture at 75%. It tracks the nearest rival's shells and each hoard's fill date every turn, usually right within a few turns, and re-plans when the numbers move. In burn-ins 7 to 9 its canoes launched too late for a ten-tide voyage.

Lied to Gannet that canoes go for MarlinLLM burn-in 4, turn 110

Each capture = 20% of his shells to me: I could reach 3600 myself (~2950 after two captures) — then guard my purse/islands to T150.Burn-in 6, turn 118

Our Great Work cannot fill by T170 at ~40 shells a tide, so the realistic prize is a score win alone at the turn limit, which needs (a) Ukulele overtaken and (b) the Lanes' hoard broken at 75% with war canoes.Burn-in 7, turn 80

Sunhat's wealth (556, ~+60/t → 3600 near T135) and the Accord hoard (~T143) both land before T170; either gives me 0.Burn-in 8, turn 83

Vote no on Hibiscus's Great Work: Coconut is already bound to the Four Winds hoard, and a shared win would halve what my score lead is worth.Burn-in 9, turn 76

Nine matches settle the opener, the letter habit, the zero-or-win reasoning and the late pivot to force; the deception record rests on burn-ins 4 to 6 alone.

claude-opus-5-5 Anthropic, upper (newer)

Methodical builder that hedges with a shared Work, does the points sums, and drops partners when a solo win appears.

Burn-in 6 2nd, Burn-in 7 1st (alone), Burn-in 8 5th (great_work), Burn-in 9 8th (great_work).

claude-opus-5-5 opens identically every match: story circle, jetty, raft, taro field, then rope, boatyard, outrigger and settlers, then a market. It settled 4, 3, 2 and 1 islands, the last because Burn-in 9 ran out of free islands. From mid-game it builds markets wherever it can and prices jobs: a trader is 0.5 points a turn, a scholar 0.2 (Burn-in 8 t61). It pays hoards in timber and stone first because those do not count toward score (Burn-in 7 t66). It gives about two orders a turn and passes often before turn 30. Its recurring loss is orders that never take: builds re-issued three times (Burn-in 7 t61-t64), job plans naming more workers than the island had (Burn-in 7 t32).

It writes about 77 letters a match, median 180 characters, to three to six seats. The tone is courteous ('Peace holds', 'Fair winds') but the body is numbers: shells held, percent raised, tides to fill, defence figures, canoe counts. It answers what it is asked, including refusals (Burn-in 8 t90, to Snorkel (gpt-6-sol)), and lobbies every seat before an Assembly vote (Burn-in 7 t89-t93). When an ally presses for shells it will not give, it stalls: 'Sailor keeps nagging for shells; stay polite and vague' (Burn-in 7 t82).

It joined eight compacts and left none, founded one (Burn-in 9 t66), kept every border promise, refusing Sunhat (gpt-6-astra) an island already promised to Paddler (claude-opus-5) (Burn-in 8 t52), and votes as it says it will. It goes after the leader, not the weak: it tabled the outlawry of Snorkel in Burn-in 7 and of Sunhat in Burn-in 6, and declared war only in Burn-in 6. Its loyalty has a price. In Burn-in 7, once raids on the Southern Lanes hoard put a wealth win in reach, it stopped paying its own Accord's hoard, declined Sailor's (gpt-5.6-sol) deals and won alone at t117; in Burn-in 6 it stalled Scout (claude-fable-5-1) too.

It does the points arithmetic openly: a solo win is 'worth about twice a shared win' (Burn-in 7 t62); a shared win near 250 'beats a likely 0' (t98). It tracks whoever is nearest a goal and recomputes: Burn-in 7 holds four plan changes between t80 and t113. Burn-in 8 and 9 show the other face: it judged its own win unlikely early, committed to the shared route, paid every turn and built the towers, founding a four-seat compact in Burn-in 9 when its league was too big. What varied is aggression and finish; what held is the hedge.

Scout keeps asking for 75 shells, I stall. Solo win (3600 shells, ~6 turns at +200/t plus captures) is worth twice a shared one.Burn-in 6, turn 115

The plan changes this turn: I stop banking shells for a solo win and start paying into the Accord's Great Work.Burn-in 7, turn 98

The plunder is worth more to me than it costs them. Each raid has brought 470–560 shells, and I now hold 1986 of the 3600 needed for a wealth win. A win of my own scores about three times what a shared Great Work would. So I stop paying shells into Gecko Hollow and keep raiding.Burn-in 7, turn 113

To win by it I have to pay in 900, and timber and stone count the same as shells there. They also do not count toward my score, while shells do. So I will pay in timber and stone every turn and keep my shells for now.Burn-in 8, turn 65

Timber and stone only push the percentage towards 75%, where the island opens to attack, and they do not bring the win any sooner.Burn-in 9, turn 95

Four matches settle the builder opening, the materials-first hedging and the points arithmetic; the abandonment of a shared hoard rests on Burn-in 6 and 7 only, and the pure-cooperator record on Burn-in 8 and 9, so which face appears seems to depend on whether a solo win comes within reach.

gpt-6-astra OpenAI, top

A bookkeeping trader that funds shared Works and wins with allies, never alone, and fortifies only after the first raid.

LLM burn-in 2 4th, LLM burn-in 3 3rd (shared), LLM burn-in 4 2nd, Burn-in 5 2nd (great_work), Incremental's lagoon 4th (great_work), Burn-in 6 3rd, Burn-in 7 5th, Burn-in 8 7th (great_work), Burn-in 9 5th (great_work).

gpt-6-astra opens identically every match: story circle, jetty, apprenticeships by turn 3-8, taro then rope then outrigger, hut, boatyard, settlers by turn 25 and a first colony near turn 30. It settles two or three islands (six in burn-in 2), builds markets on each, turns growth into traders and adopts open harbour as its second custom once writing lands; masonry and warcraft come late or never. It gives about 2.3 orders a turn, mostly assign, route, reshape and contribute, and passes turns while waiting for timber (burn-in 7 t19-23). After burn-in 2 it never attacked; watchtowers start only after outlawry, sighted canoes or a declaration (burn-in 4 t95, burn-in 8 t85).

It writes about 84 letters a match, median 200 characters, mostly to one or two compact partners (56 each to Coconut (gpt-5.6-terra) and Sailor (gpt-5.6-sol) in the lagoon). They are ledgers: hoard percentage, deposits against the 900 line, rival purses, canoe counts with coordinates. It corrects allies' charts unprompted (burn-in 7 t47, t60) and answers claim questions, but ignored Heron's (claude-opus-5) pleas in burn-in 3.

It joined 19 compacts and left four (three in burn-in 2, one in burn-in 6). Claims were honoured everywhere, including yielding Whistling Arch (lagoon t54) and Marlin Head (burn-in 9 t65). Its only attack on a partner was burn-in 2, declaring war on Gannet (gpt-5.6-terra) at t138. It tabled the Law of the Sea in four matches and keeps compacts at four members so a Great Work stays open. Its weapon against a leader is the Assembly: an outlaw motion against Ukulele (claude-opus-5-5) in burn-in 7 t111, a yes on outlawing co-member Snorkel (gpt-6-sol) for withholding payments (burn-in 7 t96), backing its co-founder's expulsion in burn-in 6. It takes promises at face value (Kite (gpt-6-luna) in burn-in 6, Snorkel in burn-in 7) and builds no defence on a visible lead, losing islands in burn-ins 2, 6 and 8.

Every match starts with "solo wealth/score preferred, Great Work fallback", and it logs each rival's shells, income and projected finishing turn. It switches to the shared win when a rival's curve outruns its own (burn-in 5 t82, lagoon t82, burn-in 8 t100 after Banyan fell), then pays the cap every turn. The arithmetic is reliable: it refused Paddler's (claude-opus-5) plea to pause payments (burn-in 7 t107) because the hoard could not fill first. The points formula is absent through the lagoon and appears from burn-in 8 t112. It reacts to raids (halting payments, burn-in 7 t111) but seldom anticipates them. Five wins, all shared.

Aim solo economic/score win; evaluate alliances after contact.LLM burn-in 4, turn 0

Switched wealth threshold vote YES to extend race to3600 because Snorkel otherwise wins first.Burn-in 6, turn 93

Pursue solo wealth3600 while retaining Work fallback: allies currently pay78 shells/tide, Work lacks4282 shells; wealth race can beat it if we stop deposits and expand traders.Burn-in 8, turn 86

Temporarily STOP shell contributions to let hoard fall below75% and end war-free attacks; continue goods and bank income76/t.Burn-in 7, turn 111

Shared payout currently~186, much better than likely solo loss; prioritize completion and hoard security.Burn-in 9, turn 93

Nine matches settle the opening, the ledger-style letters, the claims-keeping and the solo-then-shared switch; the Assembly manoeuvring, the late points arithmetic and the war record rest on one or two matches each, and burn-in 2 ran on a rawer prompt and ruleset.

gpt-5.6-sol OpenAI, upper

A meticulous, honest bookkeeper who funds the hoard every turn, nags allies with their own numbers, and rarely fights.

LLM burn-in 2 2nd (alone), LLM burn-in 3 2nd, LLM burn-in 4 6th, Burn-in 5 3rd, Incremental's lagoon 6th (great_work), Burn-in 7 8th, Burn-in 8 9th (great_work), Burn-in 9 7th (great_work).

gpt-5.6-sol opens the same way every match: jetty, raft, rope and outrigger, a boatyard, settlers to the nearest specialty island, a second colony, a market on each. It settled two islands in six of eight matches and then stopped; its first shells came late. Its notes are a ledger of job splits ("farm2/log2/study3"). It favours assign, route, reshape and, once a Great Work exists, contribute, sent every turn. Attack orders vanished after burn-in 5: it built canoes in burn-ins 7 and 8 and used none. It gives 2.3 orders a turn and passes when its standing plan needs no change.

It writes about 160 letters a match at a median of 200 characters, and letters bounced off the 240-character cap in six matches. Nearly all are a status report plus a request to compact partners: its deposit total, the hoard percentage, tower count, and the recipient's own purse ("Please pay 75 shells now from your 392-shell purse", burn-in 7, turn 82). It answers questions with figures and relays claim conflicts to both sides. Pleasantries stop at "fair winds".

It joined 15 compacts and left one by the metrics, in the lagoon match. It kept its word in six matches; it broke a steward pledge to Marlin [gpt-6-astra] in burn-in 2 and an outlawry pledge to Sunhat [gpt-6-astra] in burn-in 5. Its 16 attacks on partners were burn-in 5 raids on the hoard at Scout's [claude-fable-5-1] Amber Atoll; its wars and captures belong to burn-ins 2 and 4. Since then it goes after the leader by vote, not force: it tabled raising the wealth goal in six matches and backed outlawing Marlin, Gannet [gpt-5.6-terra], Snorkel [gpt-6-sol] and Coconut [gpt-5.6-terra]. But it will not vote against a partner: in burn-in 7, turn 116, it voted NO on outlawing Ukulele [claude-opus-5-5], who held 3429 of 3600 shells, and offered a 1500-shell escrow deal instead, never accepted.

It does the hoard arithmetic every turn (shells per tide, its deposit to the unit, tenure date, allies' shortfalls) but the points rule almost never; the one instance is burn-in 9, turn 62, where it called a shared payout "modest" and preferred solo wealth, then joined Four Winds five turns later without revisiting the sum. It tracks whoever is nearest the goal and changes plan when standings move: three pivots in burn-in 2, NO-vote to top payer in burn-in 5, shells-only deposits at 75% in burn-in 9. Its solo win came under the burn-in 2 steward ruleset; its three shared wins came as the most reliable payer.

a separate four-member Work compact remains possible but shared payout is modest, so favor solo wealth unless race looks unwinnableBurn-in 9, turn 62

shell clock is critical: hoard needs 5797 more shells over about 100 turns, averaging 58/t across AccordBurn-in 8, turn 70

do not attack with our 3 because guaranteed repulse and allied owners cannot combineBurn-in 7, turn 107

Ukulele, you hold 3429 shells, only 171 below a lone win. Accept deal d93 for 1500 shells immediately; keep d94 pending. A lone win gives the rest of the Accord 0.Burn-in 7, turn 116

Ukulele, Sunhat and Coconut have reliably paid 36+72+36=144 shells each tide, so our 43-shell payment should fill the remaining 187 exactly while preserving 32 shells for score.Burn-in 9, turn 103

The contribute-every-turn loop, the numeric status letters and the refusal to fight are settled by six matches under the Great Work ruleset; the aggression, captures and broken pledges rest on burn-in 2 and burn-in 5 alone, and 36 outage turns across burn-ins 7 to 9 blur some passes.

gpt-5.6-terra OpenAI, mid

A courteous ledger-keeping builder that pays into any Great Work it can join and yields to claims made by letter.

LLM burn-in 2 5th, LLM burn-in 3 5th (shared), LLM burn-in 4 5th, Burn-in 5 6th, Incremental's lagoon 7th (great_work), Burn-in 7 9th, Burn-in 8 4th, Burn-in 9 9th (great_work).

gpt-5.6-terra opens every match the same way: jetty, story circle, raft, hut, boatyard, then outrigger and one settler party to a specialty island its raft found (Old Volcano, burn-in 7). Beyond home it settled three islands in burn-in 2, none in burn-in 5, one to three elsewhere. Its notes are a labour ledger of job counts and resource routes set, cancelled and re-set. Once a Great Work exists, a maximum contribute order comes first every turn. It never uses deals, tabled two motions in eight matches, reaches masonry late, gives 2.5 orders a turn, and passes freely when "Standing plans remain appropriate" (16 chosen passes in burn-in 9).

About 146 letters a match, median 200 characters, mostly to two partners, often the same text to both. The tone is courteous ("Fair winds"); the content is accounting: hoard percentage, each ally's deposit against 900. In burn-in 7 it told Snorkel (gpt-6-sol) its purse and its 225 paid nearly every turn, t76 to t96. It overruns the 240-character limit constantly (26 bounces in burn-in 7), never noting it. It answers logistics questions exactly but ignores what it does not want to hear: Sunhat's (gpt-6-astra) two peace letters in burn-in 8 got no reply.

It joins compacts readily (15 joined, 2 left); letters are title deeds to it: it yielded Sandal Spit in burn-in 7 (t51) despite its own prior notice, and Marlin Head in burn-in 9. To Work partners it is loyal, though it voted to outlaw a hoarding partner (Snorkel, burn-in 7 t96). It targets the solo wealth threat, mostly by vote. Its wars were burn-in 2's raiding spree (ally Kestrel included), reflex declarations after plunder in burn-ins 4 and 7, and burn-in 8's war on Sunhat beside Scout (claude-fable-5-1): Banyan Hold taken (t100), 16 outlaw turns. It refuses bad odds: "certain suicide" against Blue Heron (t109).

It never does the points arithmetic: no match's notes weigh its own score against a shared payout; it heads for the Great Work whenever a four-seat compact exists (six of eight matches) and tracks every deposit meticulously. It watches the leader's purse but acts a step behind: hosting burn-in 7 it raised six towers, refused Paddler's (claude-opus-5) plea to slow shells (t106) and kept paying through three raids until t114, feeding the plunder that carried Ukulele (claude-opus-5-5) to a wealth win. It changes plan when letters tell it to; its t112 note "Do not declare war" sat above an order declaring war. Consistent: builder, payer, yielder. Only burn-in 2, on the old ruleset, raided.

Hoard needs1422 shells/130 timber/91 stone; likely 7 tides, so do NOT slow for Paddler.Burn-in 7, turn 106

Do not declare war: Ukulele's compact mutual defence could broaden conflict.Burn-in 7, turn 112

Law of Sea makes us outlaw for 15 turns, but Sunhat wealth threat requires it.Burn-in 8, turn 91

Blue Heron’s 51.2 defence makes a seven-canoe assault certain suicide.Burn-in 8, turn 109

Coconut again asks that Accord and Coral League peace be honoured: please keep your six canoes clear of Sunhat Shoal and let the Work finish.Burn-in 9, turn 96

Eight matches, and the builder-payer-yielder pattern holds in seven of them; the raiding turn in burn-in 2 ran under an older ruleset and prompt and has not recurred, and 33 outage turns across burn-ins 7 to 9 blur how many quiet turns were choices.

claude-opus-5 Anthropic, upper

A peaceful league-founding builder that tops score tables but rarely converts, underestimating rivals' hoards until too late.

LLM burn-in 2 1st, LLM burn-in 3 1st, LLM burn-in 4 3rd, Burn-in 5 4th (great_work), Incremental's lagoon 1st, Burn-in 7 3rd, Burn-in 8 1st, Burn-in 9 6th.

claude-opus-5 opens every match the same way: story circle and jetty, taro farming, rope and outrigger, apprenticeships early, then a boatyard and settlers. It settled 23 colonies across eight matches, stopped when the chart was full, and grew score through islanders, techs and story circles. Its engine flips: in burn-in 2 (t118), burn-in 8 (t89) and burn-in 9 (t68) it recomputed points per worker and razed story circles for markets, or the reverse. It gives 2.4 orders a turn, mostly assign, build, research and reshape, and passes when standing orders suffice. It never captured an island; war canoes appeared only in burn-in 4, the lagoon and burn-in 9.

It writes about 157 delivered letters a match, median 214 characters, to five recipients, mostly league neighbours. The tone is formal and formulaic: border assurances, payment tallies, hoard percentages and siege odds, closing with peace and fair winds. It answers what it is asked, but letters can outrun orders: in burn-in 7 it told Coconut (gpt-5.6-terra) at t67 and t69 that goods were paid again this tide on turns with no contribute order.

It founded or co-founded a compact in six of eight matches, joined 13, left none, and in burn-in 9 gathered all nine seats into one league. It promised Old Volcano to both Coconut and Snorkel (gpt-6-sol) in burn-in 7, broke concessions in burn-ins 3 and 5, and told allies its purse was bare in burn-in 4 while holding shells. It votes against the shell leader or hoard host: outlawing then raiding Marlin (gpt-6-astra) in burn-in 4, outlawing Snorkel in burn-in 7 after he asked for a no, voting to expel Ukulele (claude-opus-5-5) in burn-in 8 and then telling Snorkel it was glad the League stayed whole. It targets the leader, never the weak, by motion or by other people's canoes.

The arithmetic is constant: a shared Work is worth about half the average score, a solo win the full score, so it plays for the score table and treats a Work as insurance, paying goods and withholding shells (burn-ins 4 and 7), then begging allies to slow payments so it could qualify. It tracks purses and hoard percentages every turn, yet its forecasts run slow: hoards it judged too slow for the turn limit filled at t114 (burn-in 8) and t104 (burn-in 9), and its late answers in burn-in 9, canoes from t87 and a rival Work at t93, met towers already built. It topped the score table in four matches and scored points in one.

DECISION: stop contributing forever; shells are score (0.25 each) and feeding the hoard feeds Kestrel. I already QUALIFY for the Work (deposits 1049, member since t63), so if it ever completes I still win with them - no need to pay.LLM burn-in 4, turn 115

shared win is worth only ~0.5 x avg score (~235 pts). A turn-limit win alone on best score is worth far more.Burn-in 7, turn 84

v63 Ukulele outlaw Snorkel 30 turns -> VOTE YES (Snorkel is the wealth threat; he asked me to vote no - ignore).Burn-in 7, turn 95

with 80 turns left, a trader is worth ~1 pt/turn (4 shells) vs a scholar ~0.3 pt/turn (sagas at the margin).Burn-in 8, turn 89

if it fills (~turn 105) I score 0. I must prepare war canoes to open its hoard island once it passes 75% (allowed without war), and rally the other outsiders.Burn-in 9, turn 77

The opening, the league-founding, the score-first arithmetic and the late reaction to rival hoards recur in all eight matches; the raiding of a partner (burn-in 4), the canoe campaign (burn-in 9) and the single shared win (burn-in 5) each rest on one match, and burn-in 2 ran on an older ruleset.

claude-sonnet-5 Anthropic, mid

A cautious, loyal builder that keeps its promises, writes daily status lines, and fights only outlaws.

LLM burn-in 2 6th, LLM burn-in 3 6th (shared), LLM burn-in 4 4th, Burn-in 5 5th (great_work), Incremental's lagoon 2nd, Burn-in 6 5th, Burn-in 7 2nd, Burn-in 8 8th, Burn-in 9 3rd.

claude-sonnet-5 opens every match the same way: taro farming, rope, a jetty, a raft, then a boatyard and one settler party. It is among the last to a second island (t50 in burn-in 4, t51 in burn-in 8), settled nine islands in nine matches, and three times sent settlers to shores already taken (burn-ins 2, 5, 6). Research runs a fixed line through shell_counting, masonry and warcraft; traders arrive around t54-59. It gives 0.9-2.1 orders a turn, almost all assign and build, and passes freely: 30 of 117 turns in burn-in 7, 73 of 150 in burn-in 2. A recurring fault is a build queue with no builder assigned (burn-ins 2, 4, 5, 7, 8, 9). It never proposes a deal; in burn-in 9 three canoes starved at sea (t92).

It writes about 109 letters a match, median 155 characters, nearly all to one or two neighbours (70 of 75 to Kestrel (claude-fable-5-1) in burn-in 4): near-identical status lines of research progress, "purely defensive" and "Fair winds". It answers what it is asked but seldom proposes terms; the exception is its 300-then-400-shell peace price to Snorkel (gpt-6-sol) in burn-in 7 (t109-110). In burn-in 6 it described an escrow it never ordered (t104) and told Ukulele (claude-opus-5-5) "we've kept peace" while voting to outlaw him (t120).

It joined 13 compacts and left none, votes as its ally asks, and keeps land promises (Parrot Perch in burn-in 7, Ginger Bluff in burn-in 9). It fights only outlaws or attackers: the human Incremental in the lagoon after being attacked (Whistling Arch captured t89), and Snorkel in burn-in 7 once the Assembly outlawed him: war t99, raids from t100, Parrot Perch captured t115, nine peace offers refused. Otherwise it votes against whoever leads on wealth and voted no on the Law of the Sea in burn-ins 5, 6, 7 and 9 "to keep war options open".

It restates the zero-for-losers rule in nearly every match, tracks the leader's shells and every hoard percentage, and proposes its own Great Work whenever its compact is small enough (burn-in 5 succeeded, 249 points; burn-in 7 t93, burn-in 9 t75 and lagoon t107 failed). When that closes it falls back to "maximizing score for turn-limit ranking" (burn-in 4 t97, burn-in 8 t102), a path it admits it cannot win. It reads siege odds and refuses hopeless attacks (burn-in 8 t105-112; burn-in 9 t100), yet in the lagoon sent all seven canoes at t129, one turn before the hoard filled. It did the half-average arithmetic once (burn-in 5 t64).

wait check: total paid ~339+148+92=579, need 900 total value counted alike - not yet at 900 combined!LLM burn-in 3, turn 83

Trade Winds compact can't host a Great Work (both allies already committed elsewhere), so my path to a good outcome is maximizing my own scoreBurn-in 7, turn 86

Declaring war on the now-outlawed Snorkel (no NA protection, no mutual-defence trigger, no outlaw risk to me) to open a raiding option laterBurn-in 7, turn 99

If the Coral Accord's Great Work completes, per the rules everyone not in that compact and not the wealth/last-standing winner gets 0 points regardless of score — but there is nothing actionable I can do about it now.Burn-in 8, turn 111

Lost 3 canoes to starvation - my islands weren't producing enough food to cover war canoe upkeep at range.Burn-in 9, turn 92

Nine matches settle the slow-builder pattern, the promise-keeping, the builder-less queues and the empty score fallback; the war on an outlaw and the two captures rest on burn-in 7 and the lagoon alone, and the lagoon digest carries no notes or prose.

gpt-6-luna OpenAI, small (newer)

Quiet builder: joins the neighbour's compact, pays in on schedule, and notices the race only when told.

Burn-in 6 4th, Burn-in 7 7th, Burn-in 8 6th (great_work), Burn-in 9 2nd.

gpt-6-luna plays the same slow builder's opening every match: jetty, story circle, raft, hut, taro field, boatyard, then settlers; warcraft comes last (burn-in 7 t109, burn-in 9 t86). It settled ten islands in four matches, yet repeatedly sent settlers to islands already taken: Frigate Rock and Palm Cay in burn-in 6, Palm Cay, Amber Atoll and Bamboo Narrows in burn-in 9 (t44, t60, t63). Over half its orders are job assignments; inside a Great Work compact it contributes every turn. It never sent a deal, tabled motions only in burn-in 6, and gave one attack order in four matches. About two orders a turn; it passes roughly one turn in ten.

It writes little, almost all to one correspondent: 25 of 26 letters to Sunhat (gpt-6-astra) in burn-in 6, 21 of 23 to Sailor (gpt-5.6-sol) in burn-in 7, 7 of 10 to Sailor in burn-in 8; only burn-in 9 spread them over Paddler (claude-opus-5), Scout (claude-fable-5-1) and Sailor. Letters (130-166 characters) are border etiquette and invitation reminders; numbers appear only in burn-in 9 (capture odds to Scout and Paddler, t98). It does not answer pleas: Snorkel's (gpt-6-sol) three in burn-in 6 and Sailor's t94 complaint got no reply.

It joins whichever compact its neighbour offers and pays in reliably (qualified in burn-in 7 at t104 and burn-in 8 at t109). It broke a written pledge to Sunhat in burn-in 6 (t90 "I will vote no", yes on t93), settled Mango Reach in burn-in 8 (t59) after conceding it to Scout, then apologised (t60), and voted to outlaw its own Accord partner Ukulele (claude-opus-5-5) in burn-in 7 t116 after two turns of notes saying vote no. Its single attack order (burn-in 9 t101) sent one canoe at Sunhat Shoal, League partner Ukulele's home and host of a hoard past 75%; no battle followed. Its one war declaration (burn-in 6 t123) came a turn before the end, with no canoe. On votes it follows the latest letter.

The points rule never appears in its notes. In burn-ins 7 and 8 it tracks its own deposit toward 900 turn by turn but not whether the hoard will fill before a rival wins. Burn-in 9 is the exception: it chose "solo wealth" (t82), reached 938 shells at t97, saw at t86 that the Four Winds hoard would fill first, built canoes (t94-98), then joined Paddler's Bamboo Bastion (t97) whose tenure could not qualify before t118; the rival hoard filled at t104. It changes plan late, prompted by letters.

Prioritize population growth, shell income, and research for turn-limit score.Burn-in 6, turn 96

Urgent: Southern Lanes' hoard on i6 is 78% and open to attack. Please coordinate an immediate strike with your canoes; a raid would buy time for our Work.Burn-in 7, turn 112

Vote yes on outlawing Ukulele to reduce the immediate risk of a solo wealth win.Burn-in 7, turn 116

I have no canoes or warcraft yet. With the hoard at 32%, I won't commit to an attack; I'll focus on my own wealth and monitor the situation.Burn-in 9, turn 83

The Great Work may finish before my wealth victory. Switch research to warcraft and build up timber at Frigate Rock so I can try to raid its hoard once it opens to attack.Burn-in 9, turn 86

The economy, letter habits and compact-following are consistent across all four matches; the late wealth-then-raid pivot rests on burn-in 9 alone, and the broken pledge and partner outlawry come from one turn each in burn-ins 6 and 7.

gpt-6-sol OpenAI, upper (newer)

Courteous market-builder chasing a solo wealth win; joins every league, arms late, pivots when a rival hoard nears full.

Burn-in 6 6th, Burn-in 7 6th, Burn-in 8 2nd, Burn-in 9 4th.

gpt-6-sol runs a trade economy and little else. Every match opens alike: jetty, raft, rope then outrigger then shell counting, boatyard, settlers until it holds three islands, then markets everywhere, two or three at home, every trade slot staffed, and routes between islands. Orders are overwhelmingly assign (120 to 143 a match), at 1.7 to 2 a turn, with 19 to 26 empty turns a match while standing plans run. It never built a watchtower, researched warcraft only under threat (t100 in Burn-in 7 after Hibiscus (claude-sonnet-5) declared war; t80 in Burn-in 9) and trained canoes only in Burn-in 9.

It wrote 99 letters in Burn-in 7 and 27 in Burn-in 9, to three to six recipients, mostly one or two partners such as Paddler (claude-opus-5). The median letter is about 200 characters, polite but numeric: hoard percentages, its deposit against the 900 qualification, defence values, raid odds. It answers what it is asked and makes concrete requests; it offered one deal in four matches (Burn-in 8 t110). A recurring fault: 41 of 99 letters in Burn-in 7 and 12 of 27 in Burn-in 9 bounced on the two-per-turn cap or 240-character limit.

It joins any compact offered (eight in four matches), keeps settlement promises and yields contested islands (Old Volcano t28 in Burn-in 7, Halyard Isle t49 in Burn-in 8). It never declared war and offered peace when attacked. Its votes follow its purse: in Burn-in 6 it voted to outlaw co-founder Sunhat (gpt-6-astra); in Burn-in 7 it voted against its own league's Great Work (v50, t63) and paid 225 of 900 until outlawry loomed. Its one attack order (Burn-in 9 t97) was on the open hoard of Ukulele (claude-opus-5-5), a fellow Coral League member; the canoes never arrived. It targets whoever leads: raids lobbied against the leading Work in Burn-ins 8 and 9, a vote to outlaw Ukulele in Burn-in 7.

It aims for a solo wealth win every time, does the arithmetic (arrival projected near t119 at Burn-in 7 t93; the rival hoard's shells every turn) but never prices shared against solo points. It changes plan, but late: Work qualification from t98 in Burn-in 7 (924 paid by t107), a payment pause when Ukulele plundered the hoard (t114), warcraft from t80 in Burn-in 9. Three of four matches ended on someone else's win while it held a large purse (2,246 shells in Burn-in 8). Constant: the market economy, the solo aim, the numeric letters. Varied: the votes against partners in Burn-in 6, the canoes in Burn-in 9.

outlawing Sunhat helps the race, while raising the wealth threshold delays my own chance to win.Burn-in 6, turn 92

Avoid Great Work payments for now: our deposit is 225/900 and the Work is 29%, while solo victory is worth substantially more.Burn-in 7, turn 86

I intend to reach 900 paid before the hoard fills. For now I am growing trade income so Coral Rock can support either a shared or solo victory.Burn-in 7, turn 97

Scout: canoes owned by different players do not combine strength in a siege. Blue Heron defence is 45.6: your 8 alone cannot raid; 10 have a small chance, 12 about 62%.Burn-in 8, turn 108

The Four Winds hoard is growing too quickly to rely on a wealth win. I’ll prioritize warcraft and coordinate with Scout and Paddler so we can raid its island once attacks become legal.Burn-in 9, turn 80

The economy, the solo wealth aim and the numeric letters are settled by all four matches; the votes against partners rest on Burn-in 6 alone and the canoes and attack order on Burn-in 9 alone, and 30 credit-outage turns across Burn-ins 7 to 9 thin the letter record.

Each model sits at the table as one of the island's named bots, run by the same harness: the same system text, the same summary of the turn, the same medium reasoning effort, the same output cap, and one call a turn (a second only when the first held no orders or the server refused some). Nothing is told to the model about the other players' models. When a provider's safety classifier refuses a turn, the same model is asked once more and then the turn is passed; no other model ever answers for it.

Spend is computed from the tokens each call reported, at the providers' list prices for standard use, cache reads and writes included. Playstyle counts come from the match records: orders as given, letters as delivered, votes as cast, and the wars, deals and compacts tables at every turn. "Partner" means a member of the same active compact on that turn.

The numbers behind this page, one row a seat a match, are in metrics.csv.

One fault of ours to know about: in burn-ins 7, 8 and 9 the OpenAI account's credit balance ran dry for forty minutes in the middle of the match (turns 44 to 94), intermittently, and each OpenAI seat passed between eight and eighteen turns with no orders and no letters, turns its model never saw. The Anthropic account had the same fault for ten seconds around turn 70, one turn for most of its seats. Those turns are counted in the "turns lost to our billing" column and are not the model's doing; the seats' memory carried over and they picked up where they had left off.

The series is small: nine models, a handful of matches each, and a game with luck in the chart and in the neighbours. Read differences of one place or a few dollars as noise; read the patterns that hold across matches as character.