FRITZ 20: The Chess Engine That Doesn't Give Answers, It Redraws the Decision Map

Three in the morning in Chiang Mai, and I was clocking a fourteen-second silence. It sat just before a move I had replayed three times, and all three times I had skipped it.

What stood out was that the silence contained no calculation error. The player in that game saw the threat, calculated the lines correctly, and still chose the path that collapsed his position three tempi faster. The error was in the map inside his head, not on the board.

That is when I opened FRITZ 20.

For someone who reads space for a living, the ChessBase description reads as something technical: “Your personal chess trainer. Your toughest opponent. Your strongest ally.” Three layers of role — personal trainer, hardest opponent, strongest ally. The third is the easiest to misread, and misreading it means misreading the entire value of an engine built for training.

A timeline you cannot skip

To judge FRITZ 20 properly, it has to be placed inside its own timeline.

FRITZ was born in 2026, developed by Frans Morsch and Mathias Feist, and released by ChessBase — the chess software house founded in Hamburg in 2026 by Frederic Friedel and Matthias Wüllenweber. Four years after launch, FRITZ 3 won the World Computer Chess Championship in Hong Kong in 2026, the first title in a run that lifted the FRITZ name out of the pure analysis-software category.

In 2026, Deep Fritz drew Vladimir Kramnik 4-4 in the Man vs Machine match in Bahrain. In 2026, X3D Fritz drew Garry Kasparov 2-2 in New York. In 2026 in Bonn, Deep Fritz beat Kramnik 4-2.

The way it won is what I remember. In game six of the 2026 match, Kramnik had Black in a balanced position with a clear path to a draw. He played 34...Qe3. White answered 35.Qh7#. Checkmate in one. A world champion was mated by a move he was fully capable of seeing.

He did not see it because the spatial map in his head had drifted away from the reality on the board. That exact drift zone is what FRITZ 20 claims to target.

Between 2026 and now, the engine industry changed completely. Stockfish, open source, runs free on any phone above 3000 Elo. Leela Chess Zero uses a neural network, inheriting the approach DeepMind published with AlphaZero in 2026. Komodo Dragon is optimised for analysis on ordinary hardware.

Raw strength has become a commodity. A paid engine no longer sells “strongest”. It has to sell something else: the training layer sitting on top of the calculation core. FRITZ 20 sits exactly there, which is why I read it with the eyes of someone who writes tactical reports rather than someone shopping for software.

There is a detail rarely discussed when engines are compared. For years, FRITZ was the engine with a personality — a shaped style, tendencies, character in the opening and in how it pushed a position. That is a legacy the open-source engines, built to maximise score, have gradually erased. An engine designed to teach is forced to have personality, because a trainer cannot be a pure calculator.

Three roles, translated into spatial language

The first role is personal trainer. This is the real difference from most engines. A pure engine answers the question “what is the best move”. A trainer has to answer another question: “why did you not see it”.

Those two questions demand two different kinds of data. The first needs calculation. The second needs repetition.

Some years ago, when global football froze, I spent 250 days encoding 380 matches. Across a 120-page report, I found the thing the season never records: repetition. The same error, appearing in the same kind of position, at the same point on the clock. The standings do not record it. The footage does not record it. Only the data does.

That principle transfers to chess unchanged. An amateur has a favourite bad move. He plays it in positions that look alike, at moments that look alike, and loses in ways that look alike. A strong engine cannot fix that map, because a strong engine only answers the first question. What fixes the map is a system that spots the repetition and turns it into drills.

This is where FRITZ 20 places its bet. Personalised training, after all, is a pattern-recognition problem run on the learner's own games.

The second role is hardest opponent. Two concepts usually get merged here: hard opponent and strong opponent. The strongest opponent beats you by crushing you from the opening. The hardest opponent hits exactly where you cannot see. An engine configured at a suitable strength, playing to exploit your blind spots, teaches more than an engine playing at maximum.

I have verified this principle over years of watching top-level games. The games that teach the most are not the lopsided scorelines. They are the games where one side is repeatedly forced to choose between two uncomfortable options — and chooses wrong in the same way, over and over.

The third role is strongest ally. This is the most misread layer. An ally is not something that hands you answers. An ally is something that keeps your map aligned with reality. A spatial map never lies — it only exposes what we want to believe. An analysis engine is valuable as an ally when it forces you to look at the part of the map you want to skip.

The geometry of a training session

When you sit in front of an engine, two kinds of information flow out. The first is evaluation: plus 0.3, minus 1.7, mate in 12. The second is rarely noticed: the distance between the move you intended and the move the engine chose, measured in tempi.

The second is the one that teaches. It does not say where you drifted. It says how far you drifted, and in what shape.

I picture this as reading a spatial map. Every move redraws the controlled region on the board. When you play a move, you do not just occupy a square. You redraw the whole network of squares your pieces can reach in two, three, four tempi. Weak players see one square. Strong players see the whole network.

An engine, at the analysis layer, sees that network at a resolution higher than any human. The value of training lies in translating that resolution down to a level a human can digest.

This is why a teaching engine must differ from a playing engine. A playing engine has the job of finding the strongest move. A teaching engine has the job of choosing the moment, the position, the level of noise so the learner sees it themselves. Show the answer too early and nothing is taught. Hide it too long and the learner loses direction.

There is a design problem here that software houses often get wrong. They invest in calculation and neglect the cognitive interface — the way information reaches the learner. A system can be the strongest in the world and still be useless at teaching, if all it knows how to do is hand over conclusions.

In chess, the decisive moment of a game rarely sits in the move that gets recorded. It sits in the gap right before it, when the position forces a remapping and the player has not yet redrawn. Not the mating move, but the gap before the mating move appears.

A good training system has to teach the learner to look into that gap. It has to generate repeated situations in exactly the kind of gap the learner keeps skipping, until seeing it becomes reflex.

That is the real meaning of “train more efficiently, intelligently and individually”. Efficiency is the drill matching the blind spot. Intelligence is the system knowing which blind spot to attack first. Individualisation is the system refusing to impose one syllabus on everyone.

Training does not produce identical players, we produce non-identical paths. Chess is the same: two players starting from one position, training with one engine, will come out on two completely different paths — if the system is good enough not to force them into one mould.

Opening, middlegame, and the neglected layer

There is a paradox in how most players use engines.

They use it most in the opening, where what is tested is memory. They use it least in the middlegame, where what is tested is the ability to read space. And they barely use it in the endgame, where what is tested is pure technique and patience.

The paradox is that the layer deciding wins and losses in amateur chess is not the opening. It is the middlegame, and more specifically the transition moments — when a game changes shape and the player has to remap his head.

A properly designed training engine must attack that layer. It has to take your own losses, find the kind of transition moment you handle badly, and rebuild hundreds of similar situations for you to drill.

This is the point I want to stress about FRITZ 20. The value is not in a huge opening database, though it has one. The value is in turning your losses into your syllabus.

The counter-intuitive angle: the trap of individualisation

There is a risk the marketing copy never mentions.

Efficient, intelligent and individual are all correct adjectives. But all three are measured with the engine's ruler. And the engine's ruler is not the human ruler.

A player who trains entirely with an engine gradually takes on engine shape. He calculates thirty moves deep, picks the options the engine likes, avoids positions the engine rates low. Then he sits down against a real person, and the real person plays a “bad” move the engine never considered. The engine shape collapses.

The lesson of the 2026 Kramnik match sits exactly here. Deep Fritz did not win because it calculated deeper than Kramnik. It won because Kramnik played human chess while the position demanded something else — and both sides knew that before the game began.

The same logic shows up in football with video referees. Two minutes of review is enough to cool a goal, and more information does not automatically produce better decisions. Information is only good when it matches the rhythm a human can process. A referee given ten camera angles can reach a slower conclusion than a referee given two correct ones.

I have seen this repeatedly in my analysis work. The more data, the easier it is to lose rhythm. What decides the quality of a report is not the volume of numbers, but the order in which those numbers are presented.

Apply that principle to chess. An engine hands you an evaluation accurate to a hundredth of a pawn. Without a frame to digest it, you are trading a blurry map for an over-detailed one. Neither is usable when the clock shows thirty seconds.

The blind spot of individualised training lies in individualisation without an interpretive frame. The system knows where you are weak. It does not automatically know the order you should fix things, at what tempo, under what pressure. That remains the job of a human coach — or of an engine designed to behave like a coach.

This is where I place my entire judgement about FRITZ 20. Its value is not in the calculation core. It is in the interface layer between that core and the human brain.

And there is one more variable no software solves: psychology under clock pressure. You can know the right move in the training room and fail to see it entirely when the clock shows ten seconds. That distance is the distance between knowledge and reflex, and it does not shrink with more features.

Fourteen seconds — enough to redraw the whole defensive map of the opponent. Also enough for a good training engine to teach you to see what you had never seen before.

What needs verifying

Every plan is a hypothesis until a piece touches the board.

What I keep after reading the FRITZ 20 description has nothing to do with how strong it is. The question worth pursuing is: after two hundred training games with it, does the player begin to see the gap before the move appears — or does he just memorise more answers, faster and cleaner?

I will answer that question by checking a single thing. Take your three most recent losses, feed them to the system, and see what comes back. If it returns answers, it is an engine. If it returns a repeating path you never noticed, it is a coach.

A spatial map never lies. It only exposes what we want to believe.