Splendor looks gentle — collect some gems, buy some cards, charm a noble or two. But underneath the chips there’s a surprisingly sharp little machine, and the table arguments it starts (“should I have reserved that card?”, “are nobles a trap?”) have real answers.
So we went looking for them. We built a roster of computer players — from ones that move at random, to hand-tuned strategists, to AIs that think several moves ahead — and had them play tens of thousands of games against each other. Crucially, we always rotate who goes first, so nobody gets a turn-order freebie in the standings. Every number below comes out of those games, not out of a rulebook.
Here’s the gist before we dig in:
Does going first actually matter?
Short answer: yes, more than you’d think — and it’s the closest thing Splendor has to luck.
Here’s the clean test. Take one player, clone it, and sit the copies down against each other. Skill is now identical, so any gap in who wins is pure seat advantage. In two-player games, the player who moves first wins about 57.5% of the time. That extra tempo — one more turn to grab the card you both wanted — quietly tilts the table.
And it doesn’t wash out with a crowd — it gets bigger in relative terms. At a four-player table, a fair share of wins is 25%. The first mover takes 37.5% — half again as many wins as the seat deserves. Oh, and ties are essentially nonexistent: someone always crosses 15 first.
Bottom line — Splendor has no dice and no hidden deck, so turn order is its main “luck” axis. It’s real and it’s worth roughly a +7.5 to +12.5 point swing in win rate. House-rule it kindly: in a casual game, let the newest player go first.
So how does a game actually unfold?
Splendor is a race to 15, and the word “race” is doing real work. The winner doesn’t score the most points — they cross the finish line first, usually landing around 16 while the runner-up is stranded near 11, still mid-sentence.
Watch the points pile up round by round and you can see the shape of every game:
Two things jump out. First, nobody scores for the opening ~5 rounds. Everyone is quietly building an engine of cheap cards — points come later. Second, the winner and the runner-up look nearly identical for two-thirds of the game. The split happens late, in the conversion sprint: whoever turns their engine into points fastest pulls away and grabs 15 before the other can answer.
Bottom line — your real opponent is the clock, not the other player’s pile. Build quietly, then close hard. The lead you “feel” at point 6 means very little; the lead at point 12 is the whole game.
Why do half my cards give zero points?
Because that’s correct play — and it’s the single most counter-intuitive thing about Splendor.
A typical winner buys about 15 cards over the game, and roughly half of them are worth zero victory points. Those zero-point cards aren’t filler; they’re the engine — permanent discounts that make the expensive, point-bearing cards affordable later. Skip the engine and you simply can’t pay for the points.
Here’s the kicker: losers build a similar-sized engine — they’re not lazy. They just convert it to points more slowly and run out of turns. The popular advice to keep your engine small (“five to eight cards, don’t over-build”) doesn’t survive contact with the data; champions keep buying right to the end.
Bottom line — build the cheap engine early without guilt, then cash it in late. The edge isn’t a bigger engine than your opponent’s — it’s a faster one.
Are some gem colours secretly better?
There’s a bit of folklore about a “colour triangle” — that certain gem colours are structurally favoured by the nobles or the card costs. We checked across many fresh setups.
It’s a myth. Demand for each colour is flat — even to within about 7%, which is well inside the noise of a single game’s specific layout.
Bottom line — there’s no “best colour.” Pick your opening bonus from what this particular board rewards — which nobles are out, which cheap cards are showing — not from a colour superstition.
Should I ever reserve a card?
This is the spiciest question at the table, and the honest answer is: it depends how far ahead you’re thinking, and how many people are playing. Reserving feels powerful — you grab a card and pocket a gold wildcard — but it costs you a whole turn, and in a race to 15, turns are everything.
We pitted a bot that reserves against an identical bot that never reserves, and changed only the context:
Read it left to right. For a player who just grabs the best move in front of them, reserving is a trap — it wins only ~34% of the time, because that wasted tempo never pays off. Give the player real foresight and reserve climbs back to break-even: now it reserves on purpose (about one move in eight) and usually buys the card back. But add a full table and it slides below break-even again — at four players our deepest-thinking bot reserved its way to losing every single game. With three rivals, the card you stash gets sniped or made obsolete before your turn comes back around.
Bottom line — reserve is a scalpel, not a habit. Use it to deny a key card or to bank a needed gold — then buy. The more crowded the table, the more “just buy something” beats “reserve and wait.”
Are nobles worth chasing?
Yes — but with a light touch. Across our games, winners attract about four times as many nobles as losers (roughly 1.7 versus 0.4 per game in a full table). Those are free points, and ignoring them is a measurable mistake.
The nuance: a bot that ignores nobles loses ground, but a bot that obsesses over them loses too — contorting your buys to court a noble wrecks your tempo. And the more players at the table, the more the nobles matter: the four-player noble race is about 44% hotter than the two-player one.
Bottom line — let your engine grow into the nobles’ requirements rather than bending the whole game around them. Aim your cheap cards at the colours a noble wants, and the noble tends to arrive on its own.
So how good can a Splendor AI get?
We lined our players up on a single skill ladder (a chess-style rating — higher is stronger). It’s a tidy story of diminishing mystery:
Random play is hopeless. A bot that just grabs the nearest points is a big leap up but still soft. The hand-tuned strategists cluster into a tight band near the top — and an AI that genuinely thinks ahead climbs into that band, beating our flagship strategist about 64% of the time.
But here’s the twist we love: the reigning champion isn’t the deep-thinking one. It’s a dead-simple rule — “play the strong strategy, but never reserve.” Pure calculation reaches the top of the pack and can’t quite dethrone it. That tells you something real about the game: Splendor’s skill ceiling is mostly tempo discipline, and you don’t need a supercomputer to reach it — you need to stop wasting turns.
Is that the ceiling? Probably not. Elsewhere, a self-teaching network has reached a rating around 2068 — comfortably above our champion — so there’s headroom above the hand-crafted band. Getting there is mostly a question of raw training compute, and that race is still on. We’ll update this report when our self-taught player has something to say.
Every figure here is measured from AI-versus-AI games on Trace Studio’s Splendor engine, with seats rotated so turn order never skews the standings. We’ve kept the machinery in the background on purpose; if you want the gritty details — how the players are built, how strength is rated, how the games are sampled — those live in the project’s technical notes.