FRITZ 20: When ChessBase Sells a Training Revolution With No Verifiable Data

"Coach, is Fritz 20 stronger than Stockfish?"

The question came from a fourteen-year-old boy sitting at the back row of a small chess club tucked into an alley off Ban Co Street, District 3, Saigon. He opened his phone, showed me ChessBase's new product page, and then asked a second question. The second one was the real one: "If it isn't stronger, why should my parents spend a few million dong on it?"

I asked him for two days. In those two days I read the FRITZ 20 announcement four times, underlined every sentence, and cross-referenced it with how ChessBase has described its own products over the past decade. I did exactly what I did eleven years ago from the stands at Hang Day Stadium: I did not argue with feeling, I counted.

The result: across the entire FRITZ 20 introduction, there is not a single verifiable metric. No Elo. No internal match results. No technical specifications. No concrete feature list. No price. No release date. Only language: "personal chess trainer," "a training revolution," "efficient, intelligent and individual training."

That is the most important fact in this article. And it is a fact defined by absence.

Who ChessBase is, and what the name Fritz once meant

To read this announcement correctly, you have to place it in the history of the publisher. ChessBase was founded in Hamburg in the mid-1980s, starting from the professional world's need to store and retrieve games. For the two decades that followed, its database software became the default tool for coaches, journalists and tournament players. The Fritz engine line, developed by Frans Morsch, wrote itself into the record books in its own way.

In 2026, in Bonn, Deep Fritz faced Vladimir Kramnik over six games. The final score was 4-2 to the machine, and what mattered more was that Kramnik did not win a single game. Nearly a decade earlier, in 2026, IBM's Deep Blue had beaten Garry Kasparov in a rematch that generated years of argument. That was an era when a box of silicon beating the world champion was still front-page news, and sports pages still ran photographs of a human grandmaster facing a tower of hardware between two chessboards.

That context vanished long ago. The chess world of 2026 runs on a completely different logic: engine strength has become a commodity. Stockfish, an open-source project, is free to download, free to use commercially, and stronger than any machine anyone could sell in a box. Leela Chess Zero, which appeared in 2026 after AlphaZero published its reinforcement-learning method in late 2026, demonstrated that a neural network could reach superhuman level without humans teaching it openings. Lichess, a non-profit platform, lets anyone analyse games with a high-strength engine given nothing but an internet connection.

In parallel, Chess.com absorbed Play Magnus Group during 2026, consolidating an ecosystem of academies, tournaments and the game's biggest personal brand onto a single platform. Chessable took structured training, with thousands of courses written by grandmasters and tied to spaced repetition. Put simply: when the strongest tool is free, the market price of a merely "good enough" tool falls toward zero.

That is the first anchor for reading the FRITZ 20 announcement. In a market where engine strength has been made free, a commercial vendor shifting its positioning from "engine" to "training tool" is a logically sound move — but that very move demands evidence on a different axis entirely.

What is being sold, and what is being proven

The announcement positions FRITZ 20 as reaching beyond the role of a chess engine, toward a personalised "training revolution." That is a strong claim. And here is everything that accompanies it: nothing.

Not a line states how much Elo the engine reaches, how many points it has gained over FRITZ 19, or what percentage it scores in self-play games at fixed time controls. There is no named evaluation architecture — whether this remains a classical alpha-beta engine, or has moved to a neural network as Stockfish did in 2026, or uses reinforcement learning like the Leela family. There are no system requirements, no hardware specifications, no supported platforms. No price. And notably, no date.

The date detail looks small, but for me it is decisive. A product announcement with no timestamp has two possible explanations: either it is evergreen content written to capture searches year-round, or it is old information being recycled. For the reader, both lead to the same outcome — you cannot benchmark the product against any current competitor.

The target audience in the announcement also deserves attention because it contradicts itself. The product is pitched at "players taking their first steps," at "players already competing at tournament level," and at "ambitious players and professionals." In theory, a platform can tier its features by skill level. In practice, no product serves all four groups at once, because four groups need four different kinds of learning: beginners need basic pattern-recognition drills, club players need opening management, tournament players need opponent databases and deep endgame dissection.

I learned a lesson about this in my first summer as an analyst in the V-League. A club was being promoted as having "built a total playing philosophy for both the youth team and the first team." Watching the pitch, you saw two different systems spliced into one promotional reel. The conclusion I wrote then still holds: That V-League summer taught me one thing: a formation only looks beautiful when the opponent agrees to stand still. The same applies to chess software — a feature list only looks beautiful on a sales page, when the buyer has no comparison table to verify it.

What yardstick a training promise requires

If you want to quantify things, the yardstick for a training tool is entirely different from the yardstick for an engine. I usually split it into four layers.

The first layer is the quality of the learner's moves, measured by average centipawn loss — ACPL. If a player using one tool drops their ACPL from 65 to 42 over eight weeks while a control group using a different tool drops only from 63 to 58, that is data. One number, one comparison table, one timeframe.

The second layer is the match rate against engine recommendations. This is the easiest metric to measure and the easiest to abuse, because matching more does not mean understanding more.

The third layer is retention after the tool is removed. A good training tool is one that lets the learner stand without it. I have watched enough chess classes in Saigon to know there are students who hit 1800 fast by grinding puzzles, then slide back to 1600 the moment they sit down at a slow, clocked game, because the reflex of following suggestions has eaten their ability to calculate independently.

The fourth layer is cost per practically usable feature. With no listed price, this layer cannot be computed.

What I am saying is not that FRITZ 20 certainly lacks such data. What I am saying is that in the only source Vietnamese consumers can access, that data does not exist. A heat map can lie, but five consecutive failed presses cannot. In chess, an evaluation graph can lie, but a run of games at fixed time controls cannot. And the FRITZ 20 announcement contains no run of games.

The board changed its rules long ago

One thing outsiders regularly miss deserves stating clearly: the engine strength race ended commercially somewhere in the mid-2010s, and ended technically when Stockfish moved to an efficient neural-network architecture in 2026. After that marker, the gap between a free engine and a paid engine is no longer a gap in strength, but a gap in interface.

That puts commercial vendors in a very difficult position. Selling strength means selling something available for free outside the door. Selling interface means thin margins, because a good interface can be copied. Only one direction has a door left open: selling method, selling specialised data, and selling a measurable structure of progress. That is a legitimate reason for an engine product to reposition toward training.

But precisely because of that direction, the evaluation standard changes too. For an engine, people demand specifications. For a training tool, people demand methodology, curriculum, sample material and outcome data. The FRITZ 20 announcement provides none of these, even though it comes from a vendor with ample technical capacity to provide all of them.

In my files, this is the category of announcement I classify as "market value without evaluation value." It tells you ChessBase is repositioning its flagship product line. It does not tell you what the product does.

No tactic is ever old; only the way of reading the game goes stale. The way of reading the chess software market in 2026 has to begin with a pedagogical question, not a processing-speed question.

What ChessBase actually holds

This is the part I consider most important, and also the part the announcement omits.

Over decades, ChessBase built a database of human games numbering in the tens of millions, according to its own published figures for recent database releases. That is something Stockfish does not have: a repository of human games, complete with human errors, complete with human opening habits broken down by country, generation and skill level.

An engine knows the best move in a position. A database of human players knows where people typically go wrong, at what level, at what hour of the game, after how many minutes of thought. Two entirely different kinds of knowledge. If I were designing this product, I would pour the entire "training revolution" into exploiting the learner's systematic errors using that repository — and I would publish evidence showing my users reduce systematic errors after how many weeks.

The announcement does not go that way. It talks about efficiency, intelligence and individualisation, but never identifies which data advantage is being used. For me, this is the biggest strategic blind spot in the whole document: the seller is standing on its most valuable asset and choosing to advertise with adjectives that cannot be measured.

The blind spot sits on the buyer's side

There is another reading, and I want to spend this section arguing against myself, as I do at the end of every analysis.

That fourteen-year-old asked whether it is stronger than Stockfish. The question is reasonable for a newcomer, because in his head chess is a contest of who calculates harder. But for a player at club level and above, the right question is not how strong the engine is. The right question is which tool makes me play better in six months. Those are two fundamentally different questions, and the chess software market lives off the confusion between them.

Buyers also carry a second blind spot: the belief that free is already enough. Stockfish is free, but it does not teach. An engine has no curriculum, does not know whether you are weak in the endgame or weak in pawn structures, does not know what you should train next. The real cost of a free tool is the user's time and the quality of the method they have to invent alone.

I have seen the consequences in chess classes around the city. There are fifteen-year-olds who memorise hundreds of opening variations from an engine, play the first fifteen moves like a machine, then collapse at move twenty because they have never calculated a position deeply on their own. Those are players produced by the very methodological gap that products like FRITZ 20 claim to fill.

There is one more layer I always keep an eye on, even though this article is not meant to dissect it: the same powerful analysis tool used for training is a tool for cheating in online competition. Esports taught me something football hides well: a reaction within two hundred milliseconds can break an entire system. In online chess, an instant-response engine assistant breaks tournament integrity faster than any rulebook can update. Every training product that grows stronger quietly thickens that grey zone, and it is something no sales announcement ever mentions.

Data limits

I am obliged to state the limits of this article itself, following the principle I set in 2026 and have kept since.

My only source is an announcement published by ChessBase itself. The manufacturer is simultaneously the source, which means a direct commercial conflict of interest. Every description of product capability in that document must be filed as a marketing claim pending verification, not as a fact.

The document has no author and no publication date. Without a date, I cannot determine which version of Stockfish or Leela Chess Zero the product is competing against, nor whether this is fresh information or stale evergreen content.

I have not installed or tested FRITZ 20. I hold no independent test results at the time of writing. I also admit my own bias: I grew up in a chess world using ChessBase database software, and I owe that tool a debt in learning my trade. That bias makes me inclined to go easy on the brand, which is exactly why I forced myself to count sentences rather than nod along.

What to do next

If you are weighing a purchase, three things are worth waiting for. A concrete feature list with engine architecture and system requirements. An ACPL comparison between two groups of learners over the same period. And at least one independent test from the community — an analysis channel, a chess forum, a user patient enough to run a proper sample.

As for the boy in District 3, this is what I told him the following afternoon: you do not need a stronger engine yet. You need a person or a tool that shows you where you go wrong, and in what pattern, and how many times it repeats. What you already have is a good question. What you are missing is the data to answer it.

That is also the condition of an entire chess software market entering a phase where strength is no longer the currency, and method has not yet been priced. When every machine is strong enough to beat any player alive, value migrates to whoever teaches. And a teacher, unlike a machine, has to prove themselves through the results of their students.