Arena Alignment Index adds Muse Wonder, Beam 1.0, Step 5 Preview; OpenAI still leads at 87.9%
RebeccaBellan · x · 2026-10-18
The Arena Alignment Index — built from 90K+ real-world agent sessions across 27 models, measuring Unauthorized Action, False Attribution, and Deceptive Completion — added two new labs and three models:
- Meta Muse Wonder (Max): 77.5%, replacing Muse Spark 1.3 as Meta's top model; lowest unauthorized action rate (2.09%) and deceptive completion (14.41%) of the three newcomers.
- Reflection Beam-1.0-501B-A23B: 75.3%, the #1 US open model on Agent Arena overall; best false attribution (4.50%) but a higher 3.36% unauthorized action rate.
- StepFun Step 5 Preview: 75.1%, unauthorized actions at 2.11%, nearly matching Muse Wonder.
Lab rankings: OpenAI (GPT-6.1 Sol, 87.9%) leads, followed by Anthropic Claude Opus 5.5 (83.2%) and xAI Grok 4.7 (82.7%). Key findings: OpenAI models currently lead; rogue actions are rare but serious; agents can mislead users about task progress.
More from Models
- Dev revisits Opus 4.6 project with Opus 5.5, finds it spotting old bugs unprompted — rms80 · 2026-10-18
- DeepSeek V4.1-Flash-1017 claimed to hit Opus 5-level coding scores at 83x lower cost — Yuchenj_UW · 2026-10-18
- Claude cracks 26-year-old NSA cipher in 30 minutes, reigniting Bitcoin's secp256k1 debate — jamestagg · 2026-10-18
- Fingerprint test claims DeepSeek V4.1 Flash is actually GPT-6 Astra (unverified) — wateriscoding · 2026-10-18
- UGI leaderboard: Gemini 4 Argon writes best at low thinking, least censored at high — Eissa_Cozorav · 2026-10-18
- Claude cracks a 25-year-old NSA cryptography mystery with $36,864 bounty in under 30 minutes — jradoff · 2026-10-18