Briefs · Taipei time (GMT+8)

OpenRouter's Armor beta puts the Jev model in front of agent tool calls, blocking rule-breaking ones before they run

一句重點：Rules are written in everyday language and come in three kinds: exceptions that let a call through, things that are usually off-limits, and things that are never allowed. Using it costs nothing during the beta (Armor docs).

發生什麼事

@OpenRouter announced the Armor beta, with TypeSafe AI's Jev model making the calls. Compared with the Sonnet 5 that Claude Code uses, Jev 1.13 came out both more accurate and cheaper, according to OpenRouter, on a test set it built itself (@OpenRouter thread).

Per the docs, what separates the three kinds is whether they can be lifted. Allow stops no call by itself; its one job is to carve out room inside soft-deny. A soft-deny rule lets the call go ahead when an allow rule matches or the user asks for it explicitly. Hard-deny can't be lifted no matter what. Claude Code users who go through OpenRouter's Ori get auto mode on any model with no extra setup (Armor docs).

Correction (2026-10-22)

The original version stated that Armor did not support zero data retention, which contradicts the official docs; this has been fixed.

01Sources

  • @OpenRouter: Armor beta announcement, built on Jev, and positioned as outperforming Claude Code's auto classifier
  • @OpenRouter thread: Jev 1.13 vs. Sonnet 5 on OpenRouter's own test set
  • Armor docs: billing, what's supported (zero data retention yes; HIPAA and data-region restrictions no), latency, audit mode, Claude Code auto mode compatibility, the three rule types

More briefs →