AI Chess Coach: A Buyer's Guide to a Crowded, Confusing Field

Type "ai chess coach" into a search bar and autocomplete finishes the thought for you: free, reddit, github, app. That reddit suffix is the tell. People typing this phrase already suspect the marketing pages are lying to them, because at least fifteen products are fighting over this exact branding right now, and most of what ranks for it today is a vendor's own landing page or a programmatic content site quoting other vendors' own numbers back at them. This piece sorts the field honestly, names where the free tools already beat the paid ones, and says plainly where a small app like ours has no business claiming credit.

Five different things are being sold as one thing

"AI chess coach" gets used as a single category label for five genuinely different tools. Confusing them is most of why the roundup pages read the same and help nobody.

1. Engines. Stockfish, and anything wrapping it. An engine tells you a move was a mistake and by how much. It cannot tell you why in words you can use, or whether a human at your level would ever have found the better move. Chessify is the purest example of this category as a standalone product: it rents you cloud compute to run Stockfish and other engines faster, nothing more.
2. Human-move models. Maia, Maia-2, and now Maia-3, from the University of Toronto's CSSLab. Instead of finding the objectively best move, these are trained on millions of real human games to predict what a player at a given rating would actually play. That is the answer to "was this mistake normal for someone my rating," which an engine cannot answer at all. Lichess hosts free Maia bots, and we cover how the model itself works in a separate piece on why chess bots blunder and what Maia is.
3. LLM explainers. A chat layer sits on top of an engine's output and turns the numbers into sentences. DecodeChess, Chessvia AI ("Chessy"), Noctie.ai, and a wave of smaller apps including Sensei Chess, ChessMind AI and ChessGPT all sell this. It is also, per the research below, the least reliable layer of the five.
4. Spaced repetition and memorization tools. Chessable's MoveTrainer, FSRS, and the Woodpecker Method. These schedule when you re-see material you already own so it sticks. They do not analyze your own games or tell you anything new about your play. We go deep on this category, including where the transfer evidence is thin, in a piece on chess puzzles and what actually transfers to real games.
5. Human content libraries. ChessMood is the clearest case: GM-recorded video courses and a live community, built by titled players, with no AI anywhere in the product. It gets grouped into "AI chess coach" roundups anyway, which is exactly the category confusion this piece is trying to clear up.

Most named products blend two or three of these categories rather than sitting cleanly in one. Chess.com's own coaching features, for instance, are an engine (Stockfish-based Game Review) with an LLM-style explanation layer stacked on top at its highest paid tier, not a new category of their own. Game Review's mechanics get their own full treatment in a dedicated piece on how chess game review actually works; it appears only as one line item here.

The field in August 2026, priced honestly

Prices below are what each vendor stated at the time this was checked, 28 August 2026. Chess subscription pricing changes without much notice, and in at least two cases here the sources contradicted each other outright. Where that happened, it is stated plainly rather than hidden behind one confident number.

Three of those numbers need a flag rather than a footnote. Chess.com's own Help Center article currently lists Gold, Platinum and Diamond as three live tiers. A Chess.com contributor blog post from January 2025 says Platinum was discontinued and replaced with a "Friends and Family" plan, pricing Gold at $60 a year and Diamond at $144. Third-party 2026 roundups quote Diamond anywhere from about $29 a year up. Those cannot all be true at once, and a same-day fetch of chess.com/membership did not return a pricing table to check directly. If Chess.com pricing matters to your decision, check chess.com/membership yourself before paying, because this piece is not going to print a number it cannot stand behind. Aimchess has the same problem at smaller scale: three separate sources quote its Pro tier at $7.99, $10 and $14 a month. Treat the low end as your planning number and confirm before subscribing. Chessvia is the third: the affiliate page carrying its $7, $14 and $19 tiers has since been deactivated, and independent searches now surface $7.99, $19.99 and $29.99 instead. None of the three is safe to quote without checking the live site.

Chessigma is worth naming on its own because it advertises an "AI Supercoach" on its own blog as "the first AI agent applied to chess." Whether that feature is actually shipped today, versus still in development, could not be confirmed here. Treat it as unverified until you check the live product yourself.

Where the LLM layer actually breaks

The plain-English explanation layer is the part of "AI chess coach" that sounds most like coaching, and by the actual research, it is the least trustworthy piece of the stack.

A 2024 benchmark run across 1,000 chess puzzles found GPT-4o playing illegal moves on 12.7% of its turns. It still reached an Elo of 1,790 and solved 501 of the 1,000 puzzles, and its Elo would have topped 2,000, USCF Expert territory, if illegal moves weren't counted as losses (the original write-up calls 2,000 the national master threshold, which is wrong: national master starts at 2,200), so the underlying chess judgment was not nothing. But better than one move in eight was not a legal move at all. GPT-4 and GPT-4 Turbo Preview, tested the same way, scored 1,047 and 1,144 Elo. Most other models tested, including several Claude 3 variants, GPT-3.5 Turbo and Mistral, landed in the 100 to 500 Elo range.

That was not a one-off from an older model generation. A test run on 7 August 2025, a full year later, forced both GPT-5 and GPT-5 Thinking into an illegal move using a specific four-move opening sequence. GPT-5 Thinking held onto the illegal move even after being told directly that it was illegal. GPT-5 changed its answer but gave an incorrect rule as the justification for why. The same author traced the identical failure pattern back through GPT-3.5, GPT-4, DeepSeek-R1, o4-mini and o3, and argues it is a systemic limitation of how these models represent a chess board, not a bug that quietly disappears with the next release.

The honest read: an LLM sitting on top of a real engine's output, turning "+5.03" into a sentence about why the rook takes the pawn, is a reasonable narrator. An LLM asked to reason about the position on its own, including which moves are even legal, is not reliable, and the gap has not closed as the models got bigger. DecodeChess, the DecodeChess layer inside Chessify, and most of the LLM-explainer products in the pricing table above are doing the first, safer job: narrating a real Stockfish evaluation rather than generating chess judgment from scratch. That distinction is not visible from a landing page, and none of the vendors above advertise it either way.

The honest admission

Lichess gives away, for zero, with no ads and no upsell: unlimited Stockfish analysis, a full analysis board, unlimited Puzzle Storm, and the free Maia bots. That is the engine and the human-move model, both of them, for nothing. So no paid AI chess coach app, including ours, can honestly claim its value is having an engine or a human-move model. Lichess already has both.

If all you want is to play a human-like opponent and get a Stockfish readout after the game, Lichess does that today and you can stop reading here. The actual case for paying anyone, us included, is not access to the underlying technology. It is combining the pieces into one loop: playing against a human-move model, reviewing with the engine, and training on what the review turns up, inside one app and one account, instead of assembling the same three pieces across Lichess, a puzzle tool and a separate explainer with three logins. Chessdrive, the app we build, runs the maia3-5m checkpoint on-device as the opponent and Stockfish separately for the analysis, in that single play, review and train loop. That is a packaging claim, not a technology claim, and it is the only honest one available.

Where Chess.com and Chessable genuinely win

Two incumbents beat a small app on specific, nameable ground, and pretending otherwise would be the same dishonesty this piece is trying to avoid in the other direction.

Chess.com's Diamond tier is the only product in this list that pairs a full move-by-move Game Review with coach-style explanations on every single move plus its Insights feature, gated behind one subscription and one account, at a scale no single-founder app matches. Whatever the exact price turns out to be once you check it yourself, the depth of the structured lesson and course library behind it is real, and larger than anything a small AI-coach app ships.

Chessable's MoveTrainer is the most evidence-backed memorization system in this list, because spaced repetition scheduling for discrete material like opening lines has real cognitive-science backing, even though that evidence is imported from language-learning and general memory research rather than a chess-specific trial. If your actual problem is forgetting your own opening prep, MoveTrainer against a course you already like is a better tool for that specific job than any AI coach app on this page, ours included.

Is a chess coach worth it, and how much does one cost

Real 2026 rates, per an overview published by CircleChess, which is itself a coaching business rather than a neutral surveyor, so read the figures as a seller's own account of the market: untitled coaches run roughly $20 to $40 an hour in the US, FIDE-titled coaches (FM, IM, GM) start around $50 an hour in private settings, and advanced or high-demand coaching runs $70 to $150-plus an hour. Group classes are cheaper per student, as low as $2 to $30 an hour depending on class size, and the cited overall US average for a private lesson sits around $50 an hour. Salary-aggregator sites such as ZipRecruiter and Glassdoor also list "chess coach" pay, but that data measures employed or job-listing compensation, not freelance per-lesson pricing, and the two should not be treated as the same number.

A human coach earns that rate by doing something no software on this page, including the LLM explainers, reliably does: watching how you actually think during a game, not just what you played, and building a repertoire and a study plan around your specific weaknesses and psychology. GM Noel Studer has separately argued that Chess.com's own post-game accuracy score is not a good measure of how well someone actually performed in a game, which is a fair description of every automated score in this piece, ours included. Automated tools are consistent and cheap. A good human coach is neither, and that is exactly the trade you are paying for. If you are well under 1500 to 1800 and have not yet used the free tools seriously, that gap is closable without spending $50 an hour. Past that point, or heading into serious tournament preparation, a human coach is doing a job no app on this list actually does.

What we measured ourselves

Every other product in this list has the engine, or the human-move model, or the LLM explanation layer, and none of the ranking pages checked for this piece published a real combined number for any product, including their own. Here is one, generated by our own pipeline on 28 August 2026: maia3-js running the maia3-5m checkpoint, the smallest official Maia chess preset, for the human-move probabilities, and Stockfish 17 at depth 20 with every legal move covered, on the finish of the Opera Game, Morphy against the Duke of Brunswick and Count Isouard, Paris 1858.

At one point in that sequence, Stockfish's own top choice is 13.Rxd7, evaluated at +5.03. Morphy played it over the board in 1858, and Steinitz later called it all powerful and exact, so this is a case where a human did find the engine's own top choice. Here is the probability Maia3 assigns to that specific move, at each rating you could set it to:

Call this exactly what it is: the probability Maia3 assigns to the engine's best move in one specific position, not a move-match accuracy figure, and not the same measurement as Maia-3's published 57.1% move-matching accuracy across large sets of real human games. Those two numbers answer different questions and should not be compared to each other.

What it shows, with the caveats that come with one game and one small on-device checkpoint stated plainly: the curve is flat and slightly bumpy through club level, moving from 3.32% at 1200 up to only 4.25% at 1500, before climbing more clearly at 1800 and above. Even at 2200, four players out of five still would not play Stockfish's own top move here. An engine that shows you Rxd7 and a green +5.03 has told you the move was strong. It has told you nothing about whether you, or anyone at your rating, could plausibly have found it. That gap between correct and findable is what a coach, human or otherwise, actually exists to close, and for once it is a gap with a number attached to it instead of a marketing claim.

Which one to actually use

If your only need is a free human-like opponent and a free engine readout: Lichess, today, no account upgrade needed. If your problem is forgetting your own opening prep: Chessable's MoveTrainer against a course you like. If you want deep, move-by-move coach-style explanations at scale and don't mind the pricing being genuinely hard to pin down: Chess.com's top tier, checked directly on their site. If you are already well past what the free stack teaches and want a plan built around you specifically: a human coach, budgeted at the rates above. If what you want is the engine and the human-move model combined in one loop instead of three separate tools, that is the specific, narrow thing an app like Chessdrive is built for, and that is a packaging answer, not a technology one.

Last verified: 28 August 2026, against the sources cited above. Pricing across this field changes often, and three figures here, Chess.com's tier pricing, Aimchess's Pro pricing and Chessvia's tiers, were contradictory across sources even on the day this was checked. Confirm current pricing directly with each vendor before paying.