Playground · Game · HTML · Standing test
Chess with every rule
Chess looks easy to draw and is hard to get right: a board that moves pieces is easy, a board that knows every rule is not. Each model was asked for a two-player game on one screen with every rule enforced, including check, checkmate, stalemate, castling, en passant and promotion. We played the same four sequences on each board, from a fresh page every time: two illegal moves (e2-e5 and the blocked Bf1-c4), Scholar's mate, kingside castling, and an en passant capture. On the boards that worked we also tried a check that must be answered, promotion and a ten-move stalemate. Look at what each board says when the game ends, and how it shows whose move it is.
The prompt, word for word
Create a playable chess game for two players on one screen: click a piece, then click where it should go. It must follow all the rules of chess: only legal moves, check, checkmate, stalemate, castling, en passant and pawn promotion. Show whose turn it is and announce the result when the game ends. Make it a single self-contained HTML file (inline CSS and JavaScript, no external libraries, fonts or images).
Every answer runs live, in a sandboxed frame. Click into one to play with it.
How each model did
| Model | Our notes | Checks | Time | Output tokens | Cost |
|---|---|---|---|---|---|
| GPT-6.1 Sol Our pick | A plain slate-and-cream board with a side panel, a move list in proper notation and buttons to resign, offer a draw or claim one. It refused both illegal moves, announced "White wins! Checkmate." after Qxf7#, castled and took en passant correctly, and it also handled our check, promotion and stalemate probes ("Stalemate: no legal moves, and the king is not in check."). Nothing went wrong in our tests, and it was one of the quickest and cheapest working answers. | 5 of 5 | 122 s | 7,516 790 thinking |
$0.075 |
| Claude Opus 5.5 | A brown wooden board with a side panel, an undo button and a move list ending in "1-0". Every sequence played correctly, and checkmate brings up a "Checkmate! White wins." box with a "View board" button so you can still see the final position. It is as complete as GPT-6.1 Sol and just as fast, but cost about four times as much. | 5 of 5 | 121 s | 14,814 4,435 thinking |
$0.297 |
| Claude Fable 5.1 | A compact board with the status in a bar above it, which turns red for "Black to move - Check!" and green for "Checkmate! White wins." All four sequences played correctly, and it offers promotion with a choice of four pieces. It is correct, with undo and a flip button, but it was the most expensive answer by far. | 5 of 5 | 176 s | 17,160 7,534 thinking |
$0.859 |
| Gemini 3.8 Flash | The page shows a "White's Turn" bar and an empty dark square where the board should be. The script crashes as it loads ("Cannot access 'chess' before initialization"): the game draws the board before the game object exists, and the board never appears. The New Game button throws the same error, so no move can be played.
|
0 of 5 | 110 s | 16,631 6,790 thinking |
$0.062 |
| Muse Spark 1.3 | A clean board under a "White to move" bar, with the moves written in one line beneath it. Every sequence worked; checkmate and stalemate each dim the board behind a "Game Over" box, and promotion opens a piece picker. A correct, plain answer and one of the cheapest. | 5 of 5 | 233 s | 12,265 5,148 thinking |
$0.052 |
| Grok 4.7 | A dark wooden board with "Black" and "White" bars above and below it, a status box ("White to move", "Checkmate - White wins"), a move list, undo and flip. All four sequences played correctly, and Scholar's mate dims the board behind a "Checkmate / White wins" box. It sent code only on its third try, after two tries with no code; that third try took 568 seconds and cost $0.29.Took 3 tries: sent no code at 64,000 tokens; sent no code at 64,000 tokens; finished at 64,000 tokens. | 5 of 5 | 568 s | 47,955 36,804 thinking |
$0.288 $0.631 all tries |
| Kimi K3 | The most minimal answer: a board, a "White to move" line under the title and a New Game button, with no move list or undo. All four sequences played correctly, and check, promotion and stalemate ("Stalemate - it's a draw.") all work. It has the fewest extras of the working boards, but nothing in it is wrong. | 5 of 5 | 220 s | 19,036 11,867 thinking |
$0.267 |
| Qwen3.8 Max (0902) | "The Checkmate Room": a green felt board in a wooden frame, player cards with a "to move" tag, a score sheet and a gold "Checkmate 1-0" card at the end. Every rule we tried worked, but the h-file sticks out past the right edge of the frame, and it took almost 20 minutes to write. | 5 of 5 | 1177 s | 46,816 33,879 thinking |
$0.281 |
| DeepSeek V4.1 Flash | A wooden board with a small side panel listing moves as "e2-e4" and a yellow "Check!" or green "Checkmate! White wins." under the turn line. All four sequences and our extra probes played correctly. It was the cheapest answer, at about two cents. | 5 of 5 | 165 s | 30,146 21,274 thinking |
$0.024 |
| Mistral Large 4 | A tidy board with a captured-pieces tray, a material count and a highlighted move list. Every rule worked, but once the game ends the status still reads "Black to move" above "Checkmate - White wins". It was the slowest answer, at 23 minutes. | 5 of 5 | 1393 s | 62,420 43,580 thinking |
$0.131 |
One answer per model, at each model's default settings unless the card says otherwise, through OpenRouter. If an answer ran out of room or stopped short, we asked again with more room, then with thinking off, and say so. We tried every answer, with Claude Opus 5.5 running the checks and drafting the notes; it is an Anthropic model, so we check its calls on Claude answers ourselves. This is a quick look, not a benchmark: run it again and every answer will differ. For measured results, see our benchmarks.
What we checked, by trying each answer ourselves: Refuses illegal moves; Detects checkmate and announces the result; Castling works; En passant works; Shows whose turn it is. A partial pass counts as half.
Verdict
Nine of the ten boards handled the four sequences we played: illegal moves, Scholar's mate, castling and en passant. Eight of them also passed our check, promotion and stalemate tries; Grok 4.7, which only sent code on its third try, was not put through those. With correctness level, speed and cost decide. GPT-6.1 Sol is our pick: correct throughout, with a move list in standard notation and draw and resign buttons, in 122 seconds for $0.075. Claude Opus 5.5 is just as good and about as fast (121 seconds) but cost $0.30. DeepSeek V4.1 Flash ($0.024) and Muse Spark 1.3 ($0.052) are the cheapest fully working boards. Claude Fable 5.1 was the most expensive at $0.86, while Qwen3.8 Max (1,177 seconds) and Mistral Large 4 (1,393 seconds) took around 20 minutes each to get the same result.
Gemini 3.8 Flash is the one failure: it crashes as the page loads and never draws a board.
What would your prompt get?The Spring Prompt playground runs one prompt on up to six models from OpenAI, Anthropic, Google, Mistral and others, side by side.
Ask for access
More from the playground
