In an Oct. 24 update, LisanBench tester @scaling01 reported that Muse Wonder Full and Muse Wonder held the top two scores at maximum reasoning effort. The benchmark’s path-length leaderboard lists Muse Wonder Full (max) first at 37,152, Muse Wonder (max) second at 27,973 and Opus 5.5 (xhigh) third at 24,029. An asterisk marks the first-place run as incomplete. @scaling01 disclosed that Meta provided free access to the Muse models and said the tests and conclusions were their own.
LisanBench asks models to build long chains of non-repeating English words. Each move must have an edit distance of exactly one: a valid word can be reached by inserting, deleting or substituting a letter. The benchmark tests 50 starting words three times each and scores the valid moves. These standings measure performance on that word-chain task, not overall AI capability.
The leaderboard also lists the cost of a full benchmark run: $256 for Muse Wonder Full (max), $166 for Muse Wonder (max) and $39.60 for Opus 5.5 (xhigh). In its separate cost-efficiency table, Muse Wonder (max) scores 168 path-length points per dollar against 607 for Opus 5.5 (xhigh). The highest path-length score is not the cheapest way to get points on this test.
The rule matters, too. @scaling01 reported that removing the edit-distance-one requirement while keeping the other rules puts GPT 6 Astra medium, Gemini 4 Argon high and GPT 6.1 Sol medium at the top of an alternative scoring. That is the tester’s recalculation, not the leaderboard’s default ranking.
Comments