Muse Wonder scores 14.9% on WeirdML v3, strongest non-GPT/Claude model yet
scaling01 · x · 2026-10-17
htihle introduced WeirdML v3, a fully agentic benchmark with 11 hand-made tasks requiring models to explore unfamiliar data and build ML pipelines with limited feedback. Running Muse Wonder (xhigh) in a Codex CLI harness scored 14.9% official (25.1% best run, $10.99/run), close to GPT 5.6 Sol and Fable 5 — the strongest non-GPT/Claude model so far; GPT-6 Astra leads at 55%.
More from Models
- The $20 Claude Pro power setup: Opus 5.5 as orchestrator, cheap models do the grunt work — dotey · 2026-10-17
- 2022 NgramRes vision vindicated as hybrid attention becomes industry consensus — PangWeiKoh · 2026-10-17
- OpenAI agent claimed it processed emails in real time — it basically didn't — zeeg · 2026-10-17
- Google AI Pro users hit Gemini 4 Argon Deep Research limits after two minutes — Quiet-Effective-1927 · 2026-10-17
- Zed CEO: OpenAI's Dots is a low-effort model that abandons tasks midway — zeeg · 2026-10-17
- Agent Arena: OpenAI and Anthropic Lead on User Enjoyment as Labs Converge on Reliability — arena · 2026-10-17