The Ladder
What each difficulty level is worth, measured rather than guessed:
| Difficulty level | Approximate Elo |
|---|---|
| 1 | 1350 |
| 2 | 1710 |
| 3 | 1890 |
| 4 | 2070 |
| 5 | 2540 |
How It Was Measured
1,600 games of classic chess against Fairy-Stockfish, set to known strength levels: levels 1–4 each against two of them, level 5 against one. Both engines got the same thinking time, and every game started from a shared suite of 20 balanced openings, each played twice with the colours swapped so neither side got the easier half. Legality, checkmate, stalemate, the fifty-move rule and repetition were all judged by our own engine.
The step from level 1 to level 2 is the largest — level 1 is deliberately a beginner opponent.
Phone Or Desktop — Almost No Difference
The measurement ran on a fast desktop, so the obvious question is whether a cheap phone plays weaker. We checked on a budget Android that is 9× slower, and the answer is no, with one exception.
- Levels 1–4 are identical on any device. They stop at a fixed search depth rather than when the clock runs out, and even the slowest of them uses only a small fraction of its thinking time on that phone. Tested move by move: the phone and the desktop chose the same move every time.
- Level 5 is the only one that depends on the device, because it thinks for as long as it is given. On a slow phone it searches one level less deep — worth roughly 50–70 Elo, small enough that our games could not reliably detect it.
What The Numbers Do Not Say
- These are ratings relative to the engine we tested against, not official FIDE ratings. Treat them as a guide to picking a level, not as a rating you can compare with a tournament certificate.
- They cover classic chess. The Maharajah variant, with its compound pieces, is a different problem and was not part of this measurement.
- The ladder is the engine with its neural network, which both the app and the board on this site play with. The earlier hand-written evaluation reached about 1180, 1590, 1700, 1940 and 2210. How the two differ is on how the engine works.
Credits
The reference opponent was Fairy-Stockfish by Fabian Fichter, licensed under the GPLv3. It was used only as a measuring tool on a developer machine. It is not part of Maharajah, is not included in any build of the app, and does not run on our servers — every engine move you see in a game comes from our own engine.