Darmok has beaten humans at cryptic crosswords. Earlier this month, the computer program solved the final puzzle of the Times’s annual cryptic crossword competition about three minutes faster than Liam Hughes, 29, who became the contest’s youngest human winner.
That result marks a notable step for artificial intelligence because cryptic crosswords demand more than vocabulary, pattern matching, or brute-force search. Every clue must produce an answer that fits the grid and a wordplay explanation that works in both ordinary language and cryptic language. Computers have been circling that problem for years. Darmok has now walked straight through it.
The achievement came after a less triumphant first attempt. Darmok entered the Times championship and failed to finish, giving up on the final puzzle after 18 minutes. Its second performance was different: the program solved that same final puzzle roughly three minutes faster than Hughes.
Wordplay remains the real test
Darmok also solved a clue that defeated three of the five finalists, including Mark Goodliffe, a 14-time champion. That matters more than a clean overall time because a difficult cryptic clue can hide its answer behind puns, misleading grammar, and several layers of wordplay. A fast solver that misses the trap is still just losing with confidence.
Will Shortz, the New York Times’s puzzle editor, once framed the central question plainly: “Is the computer going to be able to solve the clues involving puns and wordplay? I don’t think so.” Darmok’s performance supplies an answer, although one successful contest does not turn every language model into a crossword champion.
Sid Chatterjee, a computer scientist at the University of Texas in Austin, described the challenge in less forgiving terms: “It is a very, very difficult problem because you have got to make a sentence that parses both in normal English and in cryptic English smoothly, and those two grammars are very, very different.” That difference is the point. Cryptic clues are not merely questions with obscure answers; they are compact machines built from competing meanings.
Darmok’s speed outside the championship adds weight to the result. It solved a Fed crossword from last week in four minutes and 10 seconds, showing that the system can handle more than one high-pressure contest puzzle. The figures attached to the account also include 45% and 42%, though no measure is assigned to either percentage.
Four models, one crossword solver
Darmok is based on four AI models: Gemini 3.1 Pro, Claude Fable 5, GPT-5, and GPT-6 Astra. Gemini 3.1 Pro and Claude Fable 5 do most of the solving, while the other models join in when needed.
That setup matters because the result does not come from a single model producing one lucky guess. Darmok combines several systems into a solving process, then applies them to a task where answers must satisfy both the clue and the grid. The machinery is less romantic than a lone genius with a pencil, but the crossword does not award points for romance.
The program joins an existing group of AI crossword efforts that includes WebCrow and Dr Fill. Alan Connor is also part of the surrounding crossword story, as is John Henderson, known as Enigmatist. Their presence places Darmok in a longer line of attempts to make machines handle puzzles built around ambiguity rather than clean instructions.
The comparison with Deep Blue, IBM supercomputer, and Garry Kasparov, chess player, is hard to avoid. Chatterjee has already raised the stakes: “If [beating humans at cryptic crosswords] was the Mount Everest challenge, I call this next one the Olympus Mons challenge.” For now, Darmok has conquered a serious peak, not the whole range.
The result was recorded under the timestamps Tue 20 Oct 2026 16.27 BST and Tue 20 Oct 2026 16.55 BST. Its importance lies less in declaring humans obsolete than in showing that AI can now manage a form of language where meaning, structure, and deliberate misdirection must work at the same time.
