Stockfish 19 vs Stockfish 18: the official notes give a number, but how do the two versions play in real games? On this page you can see both versions against the same 15 engines in CCRL test games, the difference between their gauntlets, and what a "game pair" is. Then replay three real games against Torch, move by move.
Stockfish 18 came out on 31 January 2026. Stockfish 19 followed on 5 September 2026, about seven months later. Both are free and open source. Here are the official facts side by side. For the full list of changes, see our Stockfish 19 vs Stockfish 18 comparison and Stockfish 19 Elo.
| Stockfish 19 | Stockfish 18 | |
|---|---|---|
| Released | 5 September 2026 | 31 January 2026 |
| Gain the Stockfish team measured | Up to +44 Elo vs Stockfish 18; wins more than three times as many game pairs as it loses | Up to +46 Elo vs Stockfish 17; four times as many game pairs won as lost |
| Neural network | SFNNv16: drops threat features it did not need, adds pawn-pair features | SFNNv10 with new "Threat Inputs" |
| Second (small) network | Retired | Still used (added in Stockfish 16.1) |
| Training data | Hundreds of billions of positions rescored with a strong Leela net | Over 100 billion positions of Lc0 evaluation data |
| Download | One "universal" file that finds the best code for your CPU | Separate builds per CPU type (AVX2, AVX-512 ...) |
| Licence | Free, open source (GPL v3) | Free, open source (GPL v3) |
Source: the official Stockfish 19 and Stockfish 18 release notes (GitHub and stockfishchess.org), read 12 November 2026. "Up to" is the best result of the team's own fast test games with unbalanced openings.
"Game pairs" matter here: the Stockfish team plays every opening twice, once with each colour, and counts how many of those pairs each version wins. We explain it with two real games below.
On our channel we play the two versions against each other and explain the game. Another one: New Stockfish 19 vs Stockfish 18 with genius sacrifices.
We could not find a single game of Stockfish 19 against Stockfish 18 in the CCRL, CCC and TCEC game files from 2023 to October 2026. The two versions never met there. So we did the next best thing. CCRL plays every new engine in a "gauntlet" against many others on 4 CPU cores. We took every engine that played both Stockfish 18 and Stockfish 19 with the exact same version, and counted every game.
| Opponent (same version) | Stockfish 18 | SF18 score | Stockfish 19 | SF19 score |
|---|---|---|---|---|
| Torch v4d | +0 =30 −0 | 50.0% | +0 =29 −0 | 50.0% |
| Reckless 0.9.0 | +1 =29 −0 | 51.7% | +0 =28 −0 | 50.0% |
| PlentyChess 8.0.0 | +2 =28 −2 | 50.0% | +0 =30 −0 | 50.0% |
| PlentyChess 7.0.0 | +2 =28 −0 | 53.3% | +1 =27 −0 | 51.8% |
| Obsidian 16.0 | +2 =28 −0 | 53.3% | +1 =27 −0 | 51.8% |
| Alexandria 9.0.0 | +0 =30 −0 | 50.0% | +0 =28 −0 | 50.0% |
| Viridithas 20.0.0 | +2 =28 −0 | 53.3% | +2 =28 −0 | 53.3% |
| Stormphrax 8.0.0 | +1 =29 −0 | 51.7% | +2 =26 −0 | 53.6% |
| Halogen 16.0.0 | +2 =28 −0 | 53.3% | +2 =27 −0 | 53.4% |
| Fat Fritz 2 | +4 =26 −0 | 56.7% | +5 =24 −0 | 58.6% |
| Triumviratus 6.0 | +1 =27 −0 | 51.8% | +2 =28 −0 | 53.3% |
| Astra 7.0 | +1 =29 −0 | 51.7% | +3 =27 −0 | 55.0% |
| Cinder 0.6.1 | +0 =30 −0 | 50.0% | +1 =27 −0 | 51.8% |
| Hobbes 3.0 | +1 =29 −0 | 51.7% | +1 =27 −0 | 51.8% |
| pawnocchio 2.0.1 | +1 =29 −0 | 51.7% | +1 =27 −0 | 51.8% |
| All 15 opponents | +20 =428 −2 (450 games) | 52.0% | +21 =410 −0 (431 games) | 52.4% |
Counted from the CCRL 4-CPU gauntlet games (ccrl.live broadcast archive), February - October 2026. + = Stockfish wins, = draws, − = Stockfish losses. Score = wins plus half the draws. We kept only opponents that played both versions with the exact same version number.
Look at how many draws there are: more than 90% of the games end in a draw for both versions. At the top, engines rarely lose, so a small gain in strength shows up as only a few extra wins. That is also why the two versions sit almost level on the CCRL lists, while the Stockfish team's own tests, with unbalanced openings and fast games, show a clear gap. Read chess engine ratings explained for why the numbers differ.
Each version also had its own "Stockfish 4CPU Gauntlet" on CCRL. Stockfish 18 started on its release day, 31 January 2026. Stockfish 19 started on 11 September 2026.
| Stockfish 18 | Stockfish 19 | |
|---|---|---|
| Dates | 31 Jan - 18 Feb 2026 | 11 - 25 Sep 2026 |
| Games | 697 | 577 |
| Wins / draws / losses | 106 / 587 / 4 | 49 / 527 / 1 |
| With White | +94 =253 −1 | +41 =247 −0 |
| With Black | +12 =334 −3 | +8 =280 −1 |
| Draws | 84.2% | 91.3% |
Each version's own "Stockfish 4CPU Gauntlet" on CCRL. The opponents were not the same: Stockfish 18 met older engines such as Wasp 6.50 and Fire 10, Stockfish 19 met a different, mostly newer set of engines. That is why these two columns cannot be compared directly.
Notice the colour split. Almost all wins come with White, for both versions. With Black, a top engine mostly plays for a draw against strong opponents, and gets one.
Torch is the closest rival of Stockfish (see Torch vs Stockfish). In the CCRL "Stockfish Gauntlet" (the one that is not on 4 CPUs) Stockfish 19 played 50 games against Torch v4d: 2 wins, 48 draws. Here is the first win. The comments are ours; evaluations are from Stockfish at depth 18.
What to learn: In endgames, count the weak pawns. White gave a pawn back on move 30 to split Black's pawns, then the king collected them.
The official claim for Stockfish 19 is about game pairs. A game pair is one opening played twice, with colours swapped. If you win one and draw one, you win the pair. If each side wins with White, the pair is equal. These two Stockfish 18 games against Torch v4 were played on 16 February 2026 from the same 8 book moves (the Keres Attack). Against Torch v4, Stockfish 18 scored +2 =29 −1 in that gauntlet, and this is the pair with its only loss.
Key idea: An exchange sacrifice needs lasting play. Once the queens came off, White's extra exchange decided the game.
What to learn: The same opening gave White a big edge in both games (+1.1 by our Stockfish). That is why engines play each opening twice: only the pair is fair.
The Stockfish team measured up to 44 Elo more for Stockfish 19 in its own tests, with more than three times as many game pairs won as lost. In the CCRL games above, against the same 15 opponents, Stockfish 19 scored 52.4% and Stockfish 18 52.0%, and Stockfish 19 did not lose a game. The gap is small in balanced games because almost every game is a draw.
For you as a player, both versions are far stronger than any human, and they will show the same best move in almost every position of your games. You can try both here: play Stockfish 19 online or play Stockfish 18 at full strength. To check your own games, use Stockfish 19 in our Analyser. And if you want to know what you are up against, read can you beat Stockfish 18?
On 5 September 2026, about seven months after Stockfish 18.
On 31 January 2026. It replaced Stockfish 17.1 and was itself replaced by Stockfish 19 on 5 September 2026.
The Stockfish team measured up to 44 Elo in its own tests, and Stockfish 19 won more than three times as many game pairs as it lost. In normal balanced games the gap is much smaller, because most games are draws.
Yes, a little. Against the same 15 engines in CCRL 4-CPU games, Stockfish 19 scored 52.4% with no losses and Stockfish 18 52.0% with 2 losses.
Not in CCRL, CCC or TCEC. The direct tests were run by the Stockfish team on its Fishtest network, and those are the source of the +44 Elo number.
Stockfish 19 has a new SFNNv16 neural network, drops the second small network, was trained on positions rescored by a strong Leela net, and comes as one universal download.
Stockfish 19 and Stockfish 18 are both free here, at full strength, in your browser.
