Darmok AI solves Times cryptic crossword final faster than human winner

Earlier this month, an AI model called Darmok 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. The Guardian's crossword newsletter, Cluesletter, frames it as a long-awaited milestone: almost 30 years after IBM's Deep Blue beat Garry Kasparov at chess in 1997, a program has finally outdone humans at cryptics.

The newsletter traces the history. After Deep Blue, the New York Times's puzzle editor, Will Shortz, challenged engineers: "Is the computer going to be able to solve the clues involving puns and wordplay? I don't think so." Less than a decade later, a computer named WebCrow beat 75 human crossworders at an AI conference, and by 2021 a program named Dr Fill "won" the American crossword puzzle tournament, the most prestigious of its type for American-style crosswords. Cryptics are much trickier, and as recently as 2024 the author's colleague Alan Connor asked: "Will the cryptic crossword remain the last thing AI can't conquer?"

Dr Sid Chatterjee, a computer scientist at the University of Texas in Austin, thought not, and in 2025 set out to prove it. He began work on Darmok, named after a Star Trek episode in which the crew try to talk with a species that speaks only in allegories, an experience the article likens to solving a cryptic. Darmok's first attempt at the Times championship ended in failure: it gave up on the final puzzle after 18 minutes.

This year's success came partly from solving a clue that tripped up three of the five finalists, including 14-time champion Mark Goodliffe: "Wolf trap missing its lowest crosspiece? (5)". The crossing letters gave the pattern S_A_F. SCARF and SNARF both mean "to eat greedily" (synonyms for "wolf"), but the wordplay relied on a visual element. "Trap" indicates the synonym SNARE, and removing the bottom line of the E ("missing its lowest crosspiece") turns it into an F, so the answer is SNARF. Solvers had wrongly gone for the other word. One might expect visual clues like this to be harder for an AI, but Chatterjee said he could not detect a pattern in the types of clues Darmok struggles most with. It does find puzzles by John Henderson (known to Guardian readers as Enigmatist) harder than most.

Darmok is based on four AI models: Gemini 3.1 Pro, Claude Fable 5, GPT-5 and GPT-6 Astra. The solving is primarily done by the first two, with the others joining in if needed. Chatterjee attributed the improvement this year to the extra power behind these frontier models, as well as some of his refinements, such as making Darmok wait until it had attempted all the clues before inserting answers. Readers can test themselves against it: Darmok solved a Fed crossword from last week in four minutes and 10 seconds.

The next frontier, the article says, is setting a satisfying puzzle, which will be an even bigger jump. Chatterjee: "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." He added that if beating humans at cryptic crosswords was the Mount Everest challenge, he calls this next one the Olympus Mons challenge.

Key facts

Darmok, an AI model built by Dr Sid Chatterjee of the University of Texas in Austin, solved the final puzzle of the Times's annual cryptic crossword competition about three minutes faster than human winner Liam Hughes, 29.
Darmok is based on four models: Gemini 3.1 Pro, Claude Fable 5, GPT-5 and GPT-6 Astra, with the first two doing most of the solving.
Its first attempt at the championship failed: it gave up on the final puzzle after 18 minutes. This year it cracked a clue that tripped up three of the five finalists, including 14-time champion Mark Goodliffe.
Chatterjee credits the stronger frontier models and refinements such as making Darmok attempt all clues before inserting answers.
The next challenge, per the article, is setting a satisfying cryptic puzzle, which Chatterjee calls the Olympus Mons challenge.

Why it matters

Cryptic crosswords were long treated as a holdout against AI because clues mix wordplay, puns and sometimes visual tricks. The article sets Darmok's result beside earlier milestones: Deep Blue over Kasparov at chess in 1997, WebCrow beating 75 human crossworders at an AI conference, and Dr Fill's win at the American crossword tournament in 2021. As recently as 2024, Alan Connor asked whether cryptics would remain the last thing AI can't conquer. Will Shortz had bet against computers handling puns and wordplay.

Who it affects

Cryptic crossword solvers and competitors, including champions such as Mark Goodliffe, who were tripped up by a clue Darmok solved. Setters are in the frame too: the article names setting a satisfying puzzle as the next frontier, and Chatterjee calls it a far harder problem.

How to use it

Darmok is not presented as a product or a tool anyone can use. Readers can measure themselves against it, though: the article points to a Fed crossword from the previous week that Darmok solved in four minutes and 10 seconds. The newsletter also invites readers to write a clue for the word SCARF.

How solid is it

The account is a Guardian crossword newsletter reporting a public contest result, with Chatterjee quoted on how Darmok works and why it improved. The comparison is on the final puzzle only: Darmok solved it about three minutes faster than Liam Hughes. The exact finishing times of Darmok and Liam Hughes in the final puzzle are not given, only the approximate difference. The exact date of the Times championship is not given beyond "earlier this month".

Risks and caveats

The headline result concerns one puzzle, the final. No details are given on whether Darmok competed officially, or its result against all human finalists beyond the final puzzle, and its overall standing or total score in the competition is not given. The source does not say which exact roles each of the four models plays beyond the first two doing most of the solving. No compute cost, hardware or funding for Darmok is stated. The source also gives no timescale for AI setting a satisfying cryptic puzzle.

“If [beating humans at cryptic crosswords] was the Mount Everest challenge, I call this next one the Olympus Mons challenge.”

— Dr Sid Chatterjee, computer scientist at the University of Texas in Austin, on AI setting cryptic puzzles