AI solver finishes Times cryptic final about three minutes faster than human winner

A computer program has solved a championship cryptic crossword faster than its human winner, including a visual wordplay clue that caught out several finalists. Darmok finished the Times competition’s final puzzle earlier in October 2026 about three minutes faster than winner Liam Hughes, according to the Guardian’s October 20 report.

From a wordplay challenge to a failed first attempt

The result reaches back to a challenge posed after Deep Blue beat Garry Kasparov at chess in 1997. New York Times puzzle editor Will Shortz questioned whether computers could handle crossword clues built around puns and wordplay. His skepticism concerned the interpretation of language, rather than another contest over a chessboard.

Computers subsequently made progress on American-style crosswords, which draw on general knowledge, allusion and double meanings. In 2021, a program called Dr Fill outperformed contestants in the American Crossword Puzzle Tournament. The Guardian distinguishes those puzzles from the harder cryptic challenge Darmok was built to tackle.

Sid Chatterjee, a computer scientist at the University of Texas at Austin, began developing Darmok in 2025. He named it after a Star Trek episode about communicating with a species that speaks in allegories. Its first attempt at the Times championship final ended without a solution: the program gave up after 18 minutes.

Four models—and a decision to wait

This year’s faster finish followed changes to both the models and the solving process. The Guardian identifies four models behind Darmok: Gemini 3.1 Pro, Claude Fable 5, GPT-5 and GPT-6 Astra. Gemini and Claude do most of the solving, with the other two brought in when needed.

Chatterjee attributed the improvement partly to stronger underlying models and partly to his own refinements. One change was to make Darmok attempt every clue before entering answers. The improvement was therefore not attributed solely to choosing more capable models; Chatterjee also changed when the program committed its answers to the grid.

A decisive clue asked: “Wolf trap missing its lowest crosspiece? (5)”. Letters from crossing answers supplied S_A_F. Both SCARF and SNARF can mean eating greedily, matching “wolf.” Three of the five finalists stumbled on the clue, including 14-time champion Mark Goodliffe. Darmok reached SNARF by resolving the rest of the wordplay:

1. Read “trap” as SNARE, a synonym that supplies the starting word.
2. Remove the lowest horizontal line from the E. It becomes an F, turning SNARE into SNARF.

That success does not give Chatterjee a simple map of Darmok’s remaining weaknesses. He said he could not detect a pattern in the kinds of clues it found hardest. It does, however, struggle more with puzzles by John Henderson, known to Guardian readers as Enigmatist, than with most others.

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

Sid Chatterjee, on creating cryptic puzzles, speaking to the Guardian

Writing the puzzle is the next challenge

Chatterjee sees generating a satisfying cryptic crossword as a much harder next step than solving one. His explanation points to two demands on the same sentence: it must read smoothly as ordinary English while also working as a cryptic instruction. The task he describes is not simply producing an answer, but constructing language that works on both levels.