Changelog

Changes to the models behind the Shield API. Model IDs stay stable: when a model changes, requests keep working without edits.

29 October 2026: Pay as you go

Accounts without a paid plan can turn on pay as you go at /dashboard/shield to use shield-base, shield-large and shield-tiered, billed monthly per 1 million billable tokens. See Models and plans.
shield stays free, and paid plans still include every model with no metering.
shield-base:contribute, shield-large:contribute and shield-tiered:contribute bill half price on pay as you go and keep the same redacted attack excerpts as the free tier, with the same research acknowledgement. See Contributor pricing.

shield-large now uses Shield 1.1: Shield Large's language model (L7a, Qwen3-1.7B) is retrained and served unquantized. The ensemble with SB1, the threshold and rules v2 are unchanged.
On the model card benchmark, Large's mean balanced accuracy rises from 90.4% to 91.5%, every category improves, and the share of benign texts flagged falls from 5.4% to 4.1%. That benchmark was also used to choose between training rounds; a fresh held-out test follows.
shield-tiered escalates to Large, so its escalated texts are scored by Shield 1.1. shield and shield-base are unchanged.
No API changes: same model IDs, request and response formats, limits and plans. Read the Shield 1.1 model card addendum.

Hosted Shield API with four models: shield, shield-base, shield-large and shield-tiered. Read the Shield model card.