LisanBench Major Update: Muse Wonder Takes Top Two Spots
On October 24, blogger scaling01 published a major update to the LisanBench benchmark: for the first time, it systematically tested multiple reasoning effort levels and obtained complete data across all tiers. Meta provided free access to the Muse Wonder and Muse Wonder Full models for testing, though the blogger stressed the tests and conclusions remained independent.
Confirmed
- Overall leaderboard: Muse Wonder Full and Muse Wonder took the top two spots
- Cost trend: The cost of reaching the benchmark target is dropping 22-25% per month
- Value leaderboard: Haiku 5.5, GPT 6 Luna, and DeepSeek V4.1 Flash 1017 lead on score per dollar, with Haiku 5.5 (medium) priced at $0.22
- Token efficiency leaderboard: Measured by tokens per step and tokens per point, Anthropic (Opus 5.5 low leads with 7.7 tokens per step) and the Muse Wonder family deliver the highest output efficiency
- High-cost sample: A single Fable 5.1 xhigh run cost 437 dollars
- Task mechanics: LisanBench asks models to build the longest chain of non-repeating English words, advancing each step by adding or changing letters; Muse Wonder's reasoning scaling curves are stunning
Why it matters
- This is the benchmark's first systematic comparison across multiple reasoning effort levels, filling the gap of previous single-level testing
- The value and token-efficiency slices offer complementary perspectives for model selection under different budgets and use cases
- The 22-25% monthly decline in target-reaching cost quantifies how quickly model price-performance is improving
Note that Muse Wonder access was provided free by Meta; while the blogger declares independence, readers may still want to weigh this sponsorship when evaluating the rankings.
2026-10-24 ~ 2026-10-24 · 5 related posts
Primary sources
- [source] LisanBench Mega Update: Muse Wonder Takes Top 2, Cost to Score Falls 22-25% Monthly — scaling01 · 2026-10-24
- [source] LisanBench Perf-per-Dollar: Haiku, Luna and DeepSeek Lead the Cost Table — scaling01 · 2026-10-24
- LisanBench Token Efficiency: Anthropic and Muse Wonder Lead Tokens per Step — scaling01 · 2026-10-24
- LisanBench Word-Chain Benchmark Updated: Muse Wonder Shows Strong Reasoning Scaling — scaling01 · 2026-10-24
- [source] LisanBench update: one Fable 5.1 xhigh run cost $437, Muse Wonder excels — scaling01 · 2026-10-24