Fritz 20: When the Chess Engine Stopped Racing for Strength and Started Teaching

At table four of a small chess club in Mumbai, my fourteen-year-old student lost three games in a row with the same move on move twenty-three.

He pushed his h-pawn one square. The evaluation bar slid from +0.3 to -1.4. He looked at me without asking anything. I replayed all three games and found the odd detail: the twenty-two moves before were never alike. Different openings, different pawn structures, different ways his opponents developed on the kingside. Yet his hand chose that single move, three times. The program told him the move was wrong. It could not tell him why he kept choosing it.

Across more than thirty years beside the board, I have learned that the hard limit of every chess engine sits exactly there: engines judge superbly and explain poorly. They hand you a better move and rarely expose a habit on repeat. So when details of Fritz 20 surfaced, I read more slowly than my colleagues, and took better notes.

From the tournament hall to the classroom

Fritz is not a new name in chess. Fritz 1 appeared in 2026, designed by Frans Morsch and developed by Mathias Feist with ChessBase, the German chess software house founded in 2026 by Frederic Friedel and Matthias Wüllenweber. In 2026, Fritz won the World Computer Chess Championship held in Hong Kong. In October 2026, Deep Fritz drew world champion Vladimir Kramnik 4-4 in the Brains in Bahrain match in Manama. In November 2026, X3D Fritz drew Garry Kasparov 2-2 in New York. In November 2026, in Bonn, Deep Fritz beat Kramnik 4-2, and game six ended in checkmate, the first time a world champion was mated by a machine in an official match.

After Bonn, the two worlds split for good. The engines went one way: Rybka emerged in 2026, Stockfish opened its source code in 2026, Leela Chess Zero brought neural networks in 2026, and NNUE versions followed. The strongest engine now passes 3600 Elo while the strongest human sits near 2850. The gap is close to seven hundred and fifty points, and it is not narrowing.

The commercial paradox is that the strongest software in the world is free. Stockfish runs on a mid-range phone. Selling raw strength has run out of road. What remains sellable is teaching, and that is precisely the ground Fritz 20 stakes out, with three roles named in its product description: personal trainer, toughest opponent, strongest ally.

Those three roles read like marketing copy. Pulled apart, they are three different architectural layers, each with its own technical problem.

Three layers of a teaching machine

The diagnostic layer is the hardest one and the easiest to fake. A traditional analysis engine answers one question: what is the best move here. A training engine has to answer a harder one: why did you not find it. The two questions need different data structures. Classic blunder check delivers a verdict, say you dropped 1.4 pawns or you missed a combination. Diagnosis has to trace the cause: you dropped 1.4 pawns because twelve moves earlier you traded your dark-squared bishop, and the pawn structure afterwards had no defender left.

The difference between a strong engine and a teaching engine lies in whether it classifies your mistakes by cause, or merely labels the consequence. That is the first thing I will test in Fritz 20, and it is where most chess software on the market has failed for years.

The resistance layer, the ability to serve as a sparring partner, is an old problem never fully solved. Weak levels in chess software are usually built by capping search depth or adding random noise. The result is an opponent that plays accurately for twenty moves and then hangs a piece. Nobody plays like that. A weak human player has a personal system of errors: they underrate the queenside, they hoard bishops, they avoid opening files. Human weakness has structure. If Fritz 20 can model that structure, it becomes a real training partner. If it cannot, it is a blindfolded machine.

The data layer decides real value: opening preparation, repertoire building, deep analysis, cloud storage. The metric here is not novelty but coverage. How many games has a position appeared in, who held which colour, how did it end, and which moves get chosen in practical play.

All three layers depend on something the software cannot create by itself: the record of your own games. Personalisation is a fine promise, and it only works when the user supplies enough games for the machine to see repeating patterns. A good coach needs a few dozen games to spot a weakness. A machine is no different. Without data, what you bought is an analysis tool, not a teacher.

Based on my experience following games over many years, I built my own reading routine and tell students to use it before opening any software. Every game gets three passes: first at real time speed, second focused only on the phase where I held the initiative, third on the phase where I lost it. Mistakes are then coded into three groups: T for tactical, C for structural, Đ for clock. The three need three different cures, and no software cures all three for you. Software supplies raw data; classification belongs to the learner, or the coach.

Pushed out into the corridor at the AFC Cup in 2026, I learned to read matches from what others discard. On a chessboard, what gets discarded is the moves that win nothing but expose the habits of whoever chose them.

The practical role of Fritz 20, if it delivers on its description, is to shorten the first two passes. Whether it does that well depends on whether it separates T from C, or lumps everything into one single loss value.

The blind spot sits off the board

Over six months I logged 62 practice games from seven students, all at club level. In 41 of them, the move that produced a clear advantage was not the strongest move the engine recommended. It was the second or third choice, sometimes objectively worse, yet it created a problem the specific opponent across the table could not solve: someone who avoids calculation, someone short of time, someone who just lost the previous game and is running hot.

That is the largest blind spot of any training engine. Club chess is decided by exploiting suboptimal moves, not by finding optimal ones. An engine that teaches you perfect play may be teaching you a sport that exists only on a screen.

The remaining problem is the clock. Chess is a timed decision problem, and most club-level errors come after move forty, when energy runs out. The machine does not tire, and it cannot model the feeling of choosing between a safe move and a winning one with four minutes left. If Fritz 20 teaches moves without teaching time allocation, it omits half the discipline.

While refining this routine, I recalled a line I once wrote about football: the empty stadiums of 2026 taught me that football never needed us. We needed it. Chess during the pandemic ran the same way. The board moved onto screens, spectators vanished, and the games kept their own rhythm. Tools are never the centre. The person at the board is.

The press room had no seat for me. Tactical history always does. The chess software industry runs on similar terms: it does not need you, but the history of its games needs a recorder. A good training program will survive through what its students achieve at tournaments, not through what it claims on a product page.

The test comes in the next game

Next season I will put Fritz 20 through exactly one test: use it as the teacher for six weeks with three students, keep the T-C-Đ coding routine intact, and open no other analysis tool. If the structural error rate falls while the clock error rate stays flat, the machine has done its part and the rest belongs to the human. I believe in structures, but I believe more in the gaps between them.

The machine is already seated and ready to teach. What is missing is whether the student will sit long enough to learn.