The Short Version
Quick answer: AlphaZero beat Stockfish twice. In the 2017 preview it scored 28 wins, 72 draws and no losses over 100 games. In the 2018 paper it scored 155 wins, 6 losses and 839 draws over 1,000 games. But AlphaZero was never released, and Stockfish has since added a neural network of its own. Today, Stockfish is the engine you can actually use, and it keeps winning.
That's the headline. The details are where most retellings go wrong.
I read the Wikipedia entries for both engines before writing this, because the two matches are often mixed up, and the conditions of each one matter as much as the score.
Two Very Different Engines
The match was interesting because the engines thought in completely different ways.
Stockfish, in 2017, was a classic search engine. It examined enormous numbers of positions and judged them with evaluation rules written and tuned by people.
AlphaZero came from DeepMind and learned chess from nothing but the rules. It played itself millions of times and trained a neural network to judge positions. It searched far fewer positions, but it judged each one with a learned sense of what matters.
That's why the result made headlines. It looked like a self taught program had beaten decades of human engineering in a matter of hours.
The real story is a little more complicated.
The 2017 Preview Match
The first results came out in December 2017. According to Wikipedia's AlphaZero entry, they looked like this:
- Games: 100 from the normal starting position
- Result: 28 wins, 72 draws, 0 losses for AlphaZero
- Wins by color: 25 with White, 3 with Black
- Opponent: Stockfish 8
- Time control: 1 minute per move
- AlphaZero's training: 9 hours
Nine hours of self play, then an undefeated result against the reigning engine champion. It's easy to see why it spread so fast.
The notable detail is the split by color. Almost all the wins came with White. Black's results were mostly draws.
Why People Pushed Back
Critics focused on the conditions, not the chess.
Stockfish developer Tord Romstad pointed out that a fixed one minute per move took away Stockfish's time management, which normally spends more time on critical moves. He also noted the version was a year old and running with far more search threads than it had been tested with.
Grandmaster Hikaru Nakamura put it more bluntly. He said AlphaZero was basically using a Google supercomputer while Stockfish was running on something closer to his laptop.
Neither criticism said AlphaZero was weak. They said the match didn't prove as much as the headlines claimed.
DeepMind answered by running a much larger match.
The 2018 Match Settled More
The final paper appeared in the journal Science in December 2018, with a far bigger test.
- Games: 1,000
- Result: 155 wins, 6 losses, 839 draws for AlphaZero
- Opponent: a development version of Stockfish 9
- Time control: 3 hours plus 15 seconds per move
- Hardware: both on 44 CPU cores with 32 GB hash, with AlphaZero also using 4 TPUs
This time the time control was a normal tournament one, and Stockfish had its time management back.
AlphaZero still won clearly. But Stockfish won six games, which it hadn't done in the preview. The gap was real, and smaller than the first headlines suggested.
What 839 Draws Tells You
The number that stands out in the 2018 result isn't the wins. It's the draws.
Out of 1,000 games, 839 ended level. That's about 84 percent.
It shows how close top engines already were. At that level, a strong position with correct defense is usually a draw, and wins come from small edges accumulating over many games.
So AlphaZero's result is best read as a steady edge, not a crushing superiority. It won about 15 percent of games and lost well under 1 percent.
That's still a clear result. It's just not the mismatch the word "destroyed" in old headlines implied.
Why AlphaZero Disappeared
Here's the part many people don't realize. You can't play AlphaZero.
Wikipedia notes that the actual AlphaZero program was never released to the public. DeepMind published the method, not the engine.
After the paper, AlphaZero stopped competing in chess. There was no TCEC entry, no public version and no updates.
So any claim about AlphaZero being the strongest engine today is about a program frozen in 2018. Stockfish has kept developing every year since.
That makes "who's stronger now" a different question from "who won in 2018".
Stockfish Learned the Lesson
The biggest response to AlphaZero came from Stockfish itself.
According to Wikipedia's Stockfish entry, Stockfish 12 added NNUE in September 2020. NNUE stands for efficiently updatable neural network. It's trained on millions of positions and can be evaluated quickly on an ordinary CPU.
In plain terms, Stockfish kept its deep search and added a learned evaluation, the thing that made AlphaZero special.
That combination turned out to be very strong. It gave Stockfish neural network judgment without needing special hardware.
It's also why comparing the 2018 Stockfish with today's version is like comparing two different engines.
Leela Carried AlphaZero's Idea
AlphaZero itself vanished, but its approach didn't.
Wikipedia notes that AlphaZero inspired the chess community to build Leela Chess Zero, an open source engine using the same techniques. Anyone can run it.
For a while, Leela was Stockfish's main rival at the top. The two met repeatedly in TCEC Superfinals.
Since NNUE arrived, Stockfish has pulled ahead. Wikipedia lists Superfinal wins over Leela of 53.5 to 46.5, 54.5 to 45.5, 53 to 47 and 56 to 44 in consecutive seasons.
Our Stockfish vs Leela vs Komodo comparison covers how the three engines differ in practice.
So Who Is Stronger Today?
For any engine you can actually run, the answer is Stockfish.
Wikipedia credits it with 19 TCEC titles, and it has won every TCEC season since Season 18. It's free and open source under the GPL license.
Wikipedia lists Stockfish 19 as released in September 2026. It runs on phones, laptops and servers alike.
AlphaZero can't be tested against it, because it was never released and never updated. Any head to head claim today is a guess.
The honest summary: AlphaZero won the historic matches, and Stockfish won the long race that followed.
What It Changed for Human Players
The matches did more than crown an engine. They changed how people thought about chess.
AlphaZero's games were full of long term pawn sacrifices and aggressive king walks that older engines rarely chose. Many strong players studied them for ideas.
The push of the h pawn in the middlegame became a famous theme. So did giving up material for lasting activity.
You don't need to play like an engine to learn from that. The lesson is that activity and space can be worth more than a pawn.
That idea is useful at every rating.
Using Stockfish in Your Own Games
The practical takeaway for most players is simple. Stockfish is free, and it's already inside most analysis tools.
When you review a game on the major chess sites, the engine lines usually come from Stockfish. That's the same engine family that learned from AlphaZero's example.
The skill is in how you use it. Look for the moment the evaluation swung, then ask why, before you check the best move.
Our guide on how to analyze your chess games effectively walks through a review routine that makes engine lines stick.
The engine shows the answer. You still have to learn the reason.
Common Myths About the Match
A few claims keep circulating. Here's what the sources actually say.
"Both matches used the same conditions." They didn't. The 2018 match used 3 hours plus 15 seconds per move, not 1 minute per move.
"AlphaZero never lost to Stockfish." It didn't lose in the 2017 preview. It lost six games in the 2018 match.
"AlphaZero is the strongest engine." It was never released or updated, so it can't be tested today.
"Stockfish is old technology." It added a neural network evaluation in 2020 and keeps winning TCEC.
I checked each of these against the two Wikipedia entries before listing them, since the retellings vary a lot.
Try the Engines Yourself
Reading about engine matches is fun. Playing against an engine teaches more.
You don't need AlphaZero to feel what engine strength is like. Even a limited engine punishes loose pieces and unsafe kings far faster than most humans.
Start at a level where you can win some games. Then raise it slowly. Losing every game teaches very little.
You can play a computer opponent on our play page and pick a strength that fits.
The goal isn't to beat Stockfish. It's to get a little harder to beat yourself.