What does 20 new cards a day cost you in a year?

Use Anki's simulator instead, if you can

Anki ships an FSRS simulator inside its deck options. It reads the card states in your collection, the parameters fitted to your own review history, and your real grades. That is a far better instrument than anything a web page can build, and if you have a collection you should go open it and close this tab.

This planner exists for the one case that simulator can't serve: you haven't started. There's no history to read, no parameters to fit, and the question is whether a habit of 20 new cards a day turns into 15 minutes an evening or 90 by next summer. That's a planning question, and a crude model answers it well enough to change a decision.

So the scope is narrow on purpose. Before you commit.

The planner

The model, written out

Every card walks a fixed interval ladder. Its stability starts at 3 days and multiplies by 2.5 on each success, and the interval you actually get is that stability scaled by the retention you asked for.

The scaling comes from the FSRS-6 forgetting curve, R(t) = (1 + 0.9803 × t/S)-0.1542, solved for the interval where recall probability equals your target. At 0.90 the interval equals the stability exactly, which is what that curve is built to do. At 0.95 it's 0.40 times the stability, and at 0.70 it's 9.29 times.

Then the day loop. Each day, new cards enter and get scheduled at the first rung. Every review due that day passes with your stated pass rate and lapses with the rest. A pass moves the card one rung up the ladder. A lapse sends it back to the first rung with a 1-day stability, which is roughly what Anki's relearning step does.

It runs expected values, so a batch of 10 cards at a 90% pass rate becomes 9 passes and 1 lapse rather than a coin flip. That makes it reproducible and hand-checkable, and it means the curve is smoother than a real month of study will be.

Worth doing once by hand, because it's short. At 10 new cards a day, retention 0.90 and a 90% pass rate, the first rung is 3 days. So days 1 to 3 have no reviews, day 4 has the 10 cards from day 1, and 1 of them lapses to day 5. Day 5 gets its own 10 plus that 1, which is 11. Day 6 gets 10 plus 10% of 11, so 11.1. The series converges on 10 ÷ 0.9 = 11.11 until day 12, when the first batch's second rung lands and the number steps up by 9.

No retention value here is the optimum

The retention slider runs 0.70 to 0.95, and the tool prints no recommendation, because the recommended value is an artifact of the objective rather than a property of your deck.

Running FSRS's own optimiser on one deck, one parameter set, 3 different objective functions: minimising study time gives 0.70, minimising study time per unit of knowledge gives 0.70, and maximising knowledge gives 0.95. That's the whole legal range, decided by which question you asked. Anki's shipped default is 0.90, which sits between 2 defensible ends and is the setting you get by not touching anything.

The button that used to compute a number for your own deck, Compute Minimum Recommended Retention, has been documented as removed since Anki 25.07, though the code lives on inside the simulator.

So move the slider and read the cost. That's the useful thing here, and a recommendation would be a worse answer dressed as a better one.

The 4 things it gets wrong

One ladder for every card. Real FSRS gives each card its own difficulty and stability, and stability growth depends on how retrievable the card was when you saw it. Here every card grows at 2.5x. Your easy words will run ahead of this and your leeches will fall behind.

A flat pass rate. The model uses your one number on a card you've seen once and a card you've seen 8 times. In practice the mature cards pass more often, so a real collection carries a lighter late load than this shows.

No learning steps. Anki's default learning steps put a new card in front of you at 1 minute and 10 minutes on the day you first see it. Those same-day repetitions aren't counted here, so the first-week figures are low.

Expected values throughout. There's no randomness, so there's no bad week in it. A real Monday after a skipped weekend is much worse than any point on this curve.

Given all 4, read the output as an order of magnitude. It'll tell you whether 30 new cards a day is a 20-minute habit or an hour, and it will not tell you what next Tuesday looks like.

If you want to know how much the scheduler is buying you in the first place, the benchmark everyone cites for that doesn't contain an SM-2 row at all, and the numbers have to be aggregated out of unpublished files in the repository.

Every load-bearing claim Verbamor makes is traced to its paper on the research page.