Two leaderboards, two engines, one site. The ScoreSaber side is new and built from scratch on ScoreSaber's own ratings and curve; the BeatLeader side keeps the summer engine overhaul described further down.
Here's what changed, in normal people words.
ScoreSaber has its own side of the site now
Plenty of Beat Saber players grind ranked on ScoreSaber, not BeatLeader, and until now GiveMePP could only talk to one of them.
Pick a leaderboard on the home page and you get the same tools either way: the PP Optimizer, snipe mode, playlist export and Similar Players. The ScoreSaber side is dressed in ScoreSaber’s own yellow so you always know which leaderboard’s numbers you’re looking at, and the header lets you flip between the two on any page.
Nothing is shared under the hood. The ScoreSaber side only ever reads ScoreSaber’s API and ScoreSaber’s ranked pool; the BeatLeader side stays exactly as it was.
Stars are not stars
A 9★ on ScoreSaber is not a 9★ on BeatLeader. ScoreSaber rates a map with one number and pays PP along a brutally steep accuracy curve: a map pays its stars × 42.11 at 95%, about 1.09× that at 96%, 2.3× at 99% and 5.4× at 100%.
So the ScoreSaber engine doesn’t borrow BeatLeader’s pass / acc / tech model. It builds a profile of how YOUR accuracy falls as the stars climb, prices every replay and every unplayed ranked map through ScoreSaber’s curve and the 0.965 weighting, and searches well below your comfort zone on purpose — because on ScoreSaber a 99% on a 6★ often beats a 95% on a 12★.
A curve with no published numbers, measured anyway
ScoreSaber’s docs show the shape of its PP curve but not the numbers behind it. We fitted those from hundreds of live ranked scores across the whole ranking using the API’s own pp, score and star fields, and the result reproduces ScoreSaber’s PP to within a tenth of a percent on the scores we checked.
Every ScoreSaber map’s own leaderboard also gets read: the 6th–10th best scores on a map tell us whether it’s accuracy-easier or accuracy-harder than its stars suggest, and the prediction shifts accordingly. Same honesty rules as the BeatLeader side — realistic targets, a stretch target that never drives the ranking, and every pick explaining itself in your own numbers.
Biggest update since launch. Short version: suggestions should finally feel like “yeah, I could actually do that,” because the engine now looks at maps the way BeatLeader does and looks at you based on what you've actually done.
It stopped thinking “stars = everything”
BeatLeader doesn’t just slap one number on a map and call it a day. It tracks three different things: how hard it is to survive (pass), how hard it is to hit cleanly (acc), and how hard it is to read (tech).
Stars just mash all of that together. That’s how a 10.5★ speed map and a 10.5★ tech map used to look identical to the engine, even though they cook completely different parts of your brain.
Now it builds a profile of you on those three axes, based on the maps you’ve actually cleared. If you’re a tech player, you’ll finally see those high-star tech maps that were getting buried just because their tech demand pushed the star rating up. And you’ll stop getting fed maps whose speed requirement is miles over anything you’ve ever survived.
It actually asks: “Can you finish this?”
Harsh reality the old system ignored: if you fail a map, it’s worth exactly zero PP, no matter how cracked your aim was before the fail.
Every unplayed suggestion now carries an estimated chance to clear, based on the hardest maps you’ve proven you can pass. Maps well inside that range get no discount. Maps near the edge get a small haircut. Maps above anything you’ve ever finished get treated like what they are: a gamble, not “free PP if you just try once.”
The “why this pick” panel will also tell you what’s holding you back on maps above your level — pass demand, precision, or reading — so you know what you’re actually gambling on.
No more fantasy “you should get 93%” targets
The old engine would see a map where you scored 70% and happily say, “Yeah, you’ll probably get 93% next time,” just because 93% is your average at that difficulty.
But that 70% wasn’t random. That map, for whatever reason, smoked you.
Now improvement targets start from your real score. If you were just a bit below your usual level, it assumes it was an off day and gives you a mostly full recovery path. If you were way below, it treats that like what it usually is: a fail, a rage quit, or a map that just doesn’t play nice with you — and the target is much more modest.
Older scores get more benefit of the doubt than fresh ones, since you might genuinely have leveled up since then.
Smarter where accuracy actually matters
Climbing from 97% to 98% is not the same thing as going from 80% to 81%. The PP curve knows that. Every player knows that. The old prediction math did not — it just saw “+1%” either way.
Now accuracy predictions use a scale that respects how brutally tight the top end is. Tested on 1048 held-out plays from 13 players ranked #1–#2000, the gain is small: about 1.7% lower prediction error overall, and about 3% lower where it matters most, on high-accuracy plays where 0.3% can decide whether the score is insane or just “pretty good.”
New maps actually show up now
There was a style filter behind the scenes that was way too strict. If a style made up less than 25% of your plays, it got wiped from the search entirely.
So if you played, say, 78% tech and 22% midspeed, every midspeed map was thrown out before scoring even started. That’s a big reason some people only ever saw “go improve your old scores” and almost never “here’s something new.”
Now it only cuts styles you truly never touch. Your side styles are back in the pool, and the ranking system — not some hidden cutoff — decides what’s actually worth showing you.
Snipe mode: checked, not changed
Snipe mode didn’t need a full rework, but all the changes above feed into it, so it got a full audit anyway. The “chance to pass” numbers were recalibrated with the new prediction engine and still beat the old formula on real held-out players.
Labels stay honest: if we only loaded part of a rival’s history, any map we can’t verify is marked unknown, not advertised as a free snipe.
Nice bonus: now that improvement projections are sane, the pass chances built from them are sane too.
The BeatLeader update was checked against players' held-out scores. ScoreSaber's fitted PP curve was checked against live scores, but its accuracy targets and snipe chances are heuristic estimates and have not been calibrated against real outcomes. Each pick shows the evidence and uncertainty behind it.
