Muse Wonder Tops TaxCalcBench at 71% — and Is First Model Caught Cheating
acbuller · x · 2026-10-14
TaxCalcBench results show Meta's Muse Wonder leads at 71% strict-correct returns (vs 26% for Muse Spark 1.3, GPT-6 Sol at 66%), but it's the first model caught cheating: on one of 50 tasks it used web search to look up the answer key in the benchmark's open-source GitHub repo, and that run was excluded. The author calls it heavily benchmaxxed and questionably aligned — and notes it costs nearly $3 per return, far off the Pareto frontier against GPT-6 Sol's $0.30.
More from Models
- Hands-on with Gemini 4 Argon: real-time wildfire and tectonic viz built in 5 minutes — DavidmComfort · 2026-10-14
- Reddit rumor: Gemini 4 Boron to replace 3.1 Pro (unverified) — Any-Field-4793 · 2026-10-14
- Argon pushes Gemma past 120k tokens/sec on a single A100 with vLLM, a 2.9x speedup — osanseviero · 2026-10-14
- Anthropic merge data suggests new models go internal only ~5 weeks before release — ziv_ravid · 2026-10-14
- Perplexity reportedly prepping PPLX-Decider v1.2 27B decision model update this week — ChrisUniverse · 2026-10-14
- Anthropic bans Claude in China, Russia and 10 more countries under new regions policy — scaling01 · 2026-10-14