Chess Computer: What the Term Actually Means in 2026

"Chess computer" means two different products, and the pages ranking for the exact phrase do not meaningfully explain either one. A dedicated physical device that plays chess on its own has existed since 1977. Software that plays chess gets called a chess computer too, often in the same sentence as the hardware. Checked against Bing's results for the bare term on 2026-09-06, the top results were Chess.com, Chess.com's own play page, Lichess, chess.org, a casual games portal, an educational widget page, and a FIDE-affiliated platform: seven online-play destinations, none of which address the term as a topic at all. Wikipedia's general Chess article is the closest thing to an authority among them, and its entire treatment of computer chess hardware is one sentence noting that Deep Blue beat Garry Kasparov in 1997. This page covers what none of them do: what a physical chess computer actually is, what still gets built and sold under that name in 2026, and a fact the hardware side arrived at on its own that explains why the whole topic is confusing in the first place.

The two things people mean

Software is the meaning most people land on first: an engine such as Stockfish, running inside a phone app or a website, playing against you with no physical board in the room. That side of the term is already thoroughly covered, including exactly what a difficulty setting mechanically changes and why a strength-limited engine feels wrong to play against; our companion piece on how to actually choose a computer opponent owns that ground in full, and none of it is repeated here.

Hardware is the older meaning, and on the open web, the far less explained one: a self-contained device, sensors under the squares or a screen built into the board itself, that plays chess without a phone, an account or a subscription anywhere in the loop. That is what the rest of this page is for.

When a chess computer was a box you bought

Fidelity Electronics of Chicago, founded by Sidney Samole, sold the Chess Challenger 1 starting in early 1977, generally credited as the first dedicated electronic chess computer, built by engineer Ron Nelson. Dedicated hardware got its own world championship, the World Microcomputer Chess Championship, run annually with some gaps from 1980 to 2001. Fidelity's Sensory Voice Chess Challenger won the first edition, in London in 1980, and Fidelity's programs took the first four titles outright: 1980, 1981, 1983 and 1984. The championship folded in 2001 because microprocessor power had grown enough to challenge mainframe-class hardware directly, so a separate championship for micro-based programs no longer meant anything.

Hegener and Glaser, a Munich company founded in 1969 by Manfred Hegener and Florian Glaser, launched the Mephisto brand in 1980. Under programmer Richard Lang, Mephisto machines won or tied first in seven of the microcomputer championships between 1984 and 1990: a four-way tie in 1984, a shared win in 1985, an outright win in 1986, a split first in 1987, and outright wins in both 1989 and 1990. By 1989, over 90 percent of all chess computers sold in Germany were Mephistos.

The dominance did not survive the decade. Hegener and Glaser bought its struggling American rival, Fidelity Electronics, in 1989 for roughly 7 million dollars, just as PC-based chess software began undercutting purpose-built hardware on price and, soon after, on strength. By 1992 the company had posted 28 million Deutsche Marks in losses, and in 1994 Hegener and Glaser, Mephisto brand included, was sold to Saitek for roughly 7 million Deutsche Marks. The Mephisto Montreux, released in 1995 with an estimated 500 units ever produced, was, in the words of the Chess Programming Wiki, "the last and lonesome competitor of its genre" at that year's championship. Software had already won the argument: Fritz, running on an ordinary PC, took the World Computer Chess Championship that same year, a year after Richard Lang's own program, by then called Genius and running on general-purpose hardware rather than inside a Mephisto box, had beaten Garry Kasparov in a rapid game in London.

None of this is nostalgia. It is the reason the dedicated machines still sold today are not built for grandmasters. Affordable PCs undercut dedicated hardware starting around 1990, and the trend continued as smartphones took over; per the same source, most dedicated units sold now are "of beginner and intermediate strength" and "have to compete with mobile phone chess programs" rather than with each other.

What you can actually buy in 2026

Eight products cover the current market, from standalone boards to ones that need a phone to function.

Two of these prices are worth naming as a finding rather than a footnote: they are not ordinary retailer disagreement, they are a manufacturer disagreeing with itself. Chessnut's own site showed four different Evo prices on 2026-09-06 alone: $504 on its own buyer's-guide page, $535.49 with a promo code on its product page, $629.99 on a sale listing, and $899.99 as the stated regular price, all on chessnutech.com. The ChessGenius Exclusive runs a smaller version of the same problem across two retailers: $899.95 in stock at uscfsales.com against $599 to $649, marked sold out, at chesshouse.com, a 300 dollar spread for the identical unit. The GoChess Mini AI is a different problem again: 199 dollars on the manufacturer's own store against 530 dollars in a 2026 buyer's guide, a gap this research could not resolve in either direction, so read that row as unsettled rather than as a range. Treat every figure above as a snapshot, not a settled price.

Square Off is the murkiest product in the category, and it is worth being precise about what is actually known rather than guessing either direction. Its own storefront, squareoffnow.com, and its Pro product page both 301-redirect, as of 2026-09-06, to a Miko-branded "Miko Chess Grand, Harry Potter Edition" preorder page carrying no Square Off branding at all; Miko Robotics took a majority stake in Square Off around October 2022. Separately, on a Chess.com forum thread, an owner named Rsava reported as recently as July 12, 2026, "[c]onstant disconnects, difficulty connecting/reconnecting" and "[m]oves not being registered, pieces just disappearing off the board." Those are the two facts in evidence: the brand's own storefront now routes buyers to a different product, and existing boards are still reporting connection problems well into 2026. Whether a new Square Off board can still be bought anywhere with active support is not something this page can settle either way.

DGT's own copy for the Centaur states the device "plays in a very human way for your perfect playing experience," manufacturer marketing language, not an independent measurement. The mechanism behind that claim, per DGT's own description, is an Adaptive mode with two settings, friendly and challenging, that gets weaker after a loss and stronger after a win, tracking which openings a player knows along the way. On the HIARCS Chess Forums, a user named delb0y described the result in terms DGT would probably be glad to quote: "Feels like playing across the board against a slightly better opponent, who's happy to play anytime." Other users on the same thread were less impressed, reporting the Centaur simply outplaying them rather than easing off, and one, kiwimotard, raised a complaint worth taking seriously: the Centaur "can offer gifts (free pieces)" to keep a losing player in the game. That is a cruder mechanism than predicting what a human would actually play; handing over a piece on purpose is a strength adjustment bolted onto a full-strength engine, not a model of how a person actually errs.

How strong is one, really, and who is still checking

For decades there was a real, independent answer to "how strong is this chess computer": the Swedish Chess Computer Association, SSDF, founded in 1984, tested both dedicated hardware and PC software on standardized hardware at fixed tournament time controls and published error margins alongside every rating. An archived SSDF list dated May 24, 1997 puts the strongest dedicated-era machines at Mephisto Genius 2.0, rated 2339, and Mephisto Gideon Pro, rated 2272, both running on 486-class hardware modules. That is roughly where the best dedicated hardware of that era peaked under independent testing: about 2300 to 2340 Elo.

Compare that with where the strongest software sits today. As of early 2026, CCRL's main list puts Stockfish at roughly 3653 at a 40/15 time control, with an estimated real-hardware performance rating in the 3970 to 4060 range; a separate CCRL sub-list from late December 2025 has Stockfish 17.1 at 4081. Call it roughly 3650 to 4080 depending on the exact list and time control. The gap between the best dedicated hardware of 1997 and the best software of 2026, both independently measured where a number exists at all, runs somewhere between 1,300 and 1,700 Elo points.

Whether anyone still measures the hardware side at all is the real finding here, and it deserves care rather than confidence. Wikipedia's page on the SSDF states the list was discontinued on December 31, 2023, but that sentence carries no citation, and a directly checked archive of the actual final 2023 list only calls it "the final list of 2023," without announcing an end to the project anywhere in the thread. So: reported, not confirmed. What is confirmed is that no current, independent test of any 2026-era dedicated chess computer's playing strength turned up anywhere in this research. Every Elo figure attached to a hardware product sold today is the manufacturer's own claim. Millennium states 2350 to 2400 Elo for the ChessGenius Exclusive. Miko states "over 3380" for the engine in its Chess Grand line and calls it "the same world-champion-level engine used by grandmasters to train." Both are interested parties selling the product the number describes, and neither claim has been independently audited.

The vacuum left behind has been filled by exactly what you would expect. Several 2026 buyer's guides ranking the "best chess computer" are wholly or heavily monetised through Amazon affiliate links, and they print Elo numbers anyway: 2300 for a Vonset L6, 2200 and up for an APMCETL LC2B. One of those pages says the quiet part itself, conceding there is "no independent verification of ELO ratings" in the very category it is ranking. They do not even agree on which products exist, chesswatch.com's list omitting the Chessnut Evo entirely while boundbyflame.com features it prominently. When every number available comes either from the seller or from a page earning a commission on the sale, the honest report is that nobody currently knows how strong these machines are.

The board that ships Maia

One product from the table above is the reason this page exists at all. The Chessnut Evo, a standalone board with a 12.3-inch touchscreen and its own 8-core Android system, ships two selectable engines rather than one: Stockfish and Maia. Per Chessnut's own FAQ, Maia's rating range on the Evo runs from 1100 to 1900, with Stockfish covering roughly 600 to 3190 on the same hardware. Marketing copy for the board states plainly that Maia "plays like a human, not a machine," and elsewhere describes it, in the manufacturer's own words, as "the only chess engine trained on millions of human games" among boards in its category, a comparative claim worth reading as exactly that: Chessnut's own positioning against its competitors, not an independently verified finding.

What real owners say about the difference is more convincing than either engine's spec sheet. On a Chess.com forum thread specifically about the Evo, a user named jonthescone wrote: "I enjoy playing Maia more than Stockfish. Stockfish seems to make the occasional silly move...whereas Maia just seems to pick maybe the second or third best move," adding that "it is very hard to distinguish Maia's play from human play." Another user, nik1111, put the mechanism plainly: "Maia is trained on more than 2 millions of human games and that's why her errors are also human." A third, agatti1970, drew the contrast against strength-first engines directly: "Maia is specifically built to NOT play the best of the best...created initially to play at lower strength, in a human way."

A company selling a $500 to $900 physical board does not add a second engine for no reason. It added Maia because Stockfish, simply turned down, does not produce the kind of opponent these owners describe wanting: one whose mistakes look like a person's mistakes rather than a strong player's mistakes with the volume lowered. Our sister page works through that exact mechanism in full, covering what a difficulty setting mechanically changes and why a strength-limited engine feels wrong; see how to actually choose a computer opponent, and our own explainer on how the Maia chess model works for the model itself. A currently sold competitor's hardware arriving at the same conclusion Chessdrive is built around is not something a page selling that hardware will print. We are not selling it, so we can.

What "plays like a human" looks like in numbers

Here is one number a difficulty slider cannot produce, because it does not move in a single direction. We ran our own maia3-5m checkpoint against Scholar's Mate, 1.e4 e5 2.Bc4 Nc6 3.Qh5 Nf6?? 4.Qxf7#, across rating bands from 600 to 2200, asking one question: when the mate in one is sitting on the board, does the model at each rating actually deliver it? At 600, it does 86.8 percent of the time. At 1200, a full 600 rating points higher, that drops to 80.6 percent, meaning the stronger band is less likely to deliver the mate than the weaker one. No engine's strength setting produces a curve that runs backwards like that, because turning a skill dial down only ever makes an engine worse in one direction; a model trained on human games is reproducing an actual population of players, wrong turns included, and a population does not move in a straight line. The full ten-band dataset behind that number, including where the curve turns back upward and why, lives in our complete taxonomy of what chess AI actually covers; we do not repeat the rest of it here.

So should you buy one

Mostly, no. A phone already in your pocket runs Stockfish for free, at a strength far past anything a dedicated machine has ever been independently measured at, and Lichess specifically carries no paid tier anywhere in the product. Nothing on this page changes that fact.

Buy a physical board if the thing you actually want is the board: no screen beyond the one built into it, no account, no notification pulling you back to a phone mid-game. Once you accept that framing, you are not paying several hundred dollars for the engine inside it, since the same engine, or a stronger one, runs free on a phone; you are paying for the wood, the sensors and the fact that it is furniture rather than an app. On that basis: the DGT Centaur is the most honestly positioned of the group, since DGT does not even publish an Elo claim, unusual among its competitors and arguably more honest, even though it makes comparison harder. The Millennium ChessGenius line is for someone who wants Richard Lang's decades-old program in a dedicated box rather than on a phone screen. The Chessnut Evo is for someone who wants the two-engine idea, Stockfish and Maia both, built into furniture instead of an app. Square Off is not a purchase to make right now with any real confidence about what you would be buying, or from whom.