Gambit ROI: Which Gambits Beat Their Engine Evaluation
Across 587 chess gambits with at least 50,000 Lichess games each, the Queen's Gambit Accepted: Normal Variation: Nc6 has the best practical-vs-engine gap — 69.0% for White against a Stockfish evaluation of only +0.87 — while the King's Gambit Declined: Falkbeer Countergambit: fxe5 is the worst, scoring just 14.5% for White despite Stockfish only rating the position -4.49.
A gambit's whole pitch is trading objective material or structure for practical chances. This study puts a number on that trade: an ROI score of White win % − 50 − (Stockfish eval in pawns × 100). A positive score means the gambit scores better for White in real games than its engine evaluation implies — the 'surprise value' that makes a gambit dangerous at the board even when it's not dangerous on a chess engine's assessment. A negative score means the opposite: the position is punished in practice roughly as hard as, or harder than, the engine already predicts.
The gambits with the highest surprise value
Five lines lead the ROI ranking: the Queen's Gambit Accepted: Normal Variation: Nc6 (+18.13), Queen's Gambit Accepted: Old Variation: b5 (+16.20), Center Game: von der Lasa Gambit: Bc5 (+14.94), and the Vienna Gambit: exf4 in both its Vienna Game and standalone forms (+13.54 and +13.53). These are lines where the practical result at 1200–1600 rating is meaningfully better for the gambiteer than a near-equal or only-slightly-worse engine score would suggest — the opponent's defensive task is harder than the evaluation implies.
The gambits that underperform their evaluation
At the other end, the King's Gambit Declined: Falkbeer Countergambit: fxe5 (-31.01), Bird Opening: Dutch Variation, Dudweiler Gambit (-27.90), and Queen's Pawn Game: Zurich Gambit (-18.84) score even worse in practice than their already-unfavorable Stockfish evaluations predict. These are gambits where the compensation genuinely fails to materialize against real opponents, not just against engines.
Gambit ROI ranking
Download the data
Free to reuse under CC BY 4.0 — please link back to this page.
gambit-roi.csv gambit-roi.jsonMethodology: win rates and game counts from the Lichess Opening Explorer (blitz + rapid games, 1200–1600 rating bands); position evaluations from Stockfish at depth 16. Snapshot: July 2026. See the full dataset methodology for details.