ZeroDrift Anchor 3.0: enforcement SLMs for agent messages vs FINRA/SEC-style rules

ZeroDrift put Anchor 3.0 into general availability on 2026-09-23: a family of small language models that check AI-generated communications against regulations and company policy before send, via ZeroDrift’s Enforcement API (product, GlobeNewswire).

This post draws on ZeroDrift’s product page, the vendor PR wire and one secondary report. The headline numbers below are vendor-run / self-reported: not an independent certification, and not a FINRA or SEC endorsement.

What it returns

Per ZeroDrift, Anchor returns a verdict, the rule cited, the exact passage, and (on rewrite tiers) a verified rewrite. It sits on agents, email, chat, voice, marketing, documents, and APIs — cloud or VPC. Runtime posture: a small model post-trained for enforcement beside a deterministic rules engine; every rewrite is re-checked before ship.

Sizes and list prices

One enforcement = one message up to 2,000 tokens; rewrite + verification included where applicable; output tokens not billed. New accounts get $10 free credits (product):

Tier	Role	Base	List price / enforcement
Anchor 3.0-mini	Classify / whole-message verdict (no rewrite)	Gemma E4B	$0.002
Anchor 3.0 (flagship)	Per-line find + rewrite + verify; LoRA custom policies	Gemma E4B	$0.01
Anchor 3.0-max	Reasoning / upload policies at call time	Qwen3.8-27B	$0.05

GlobeNewswire frames Mini + flagship as ~9B/4B active and Max as 27B.

Regulatory packs (vendor claim)

ZeroDrift says packs for US securities (SEC · FINRA · Reg BI), banking/lending, insurance, CFTC/NFA, fair housing, marketing channels (TCPA/CAN-SPAM/FTC), EU/UK/APAC promotions, digital assets, cyber/privacy, and EU AI Act / NIST AI RMF live in the weights — activate subsets in Policy. That is product framing for communications enforcement, not a claim that regulators certified the models.

Vendor-run FINRA numbers (label clearly)

ZeroDrift’s methodology line: “FINRA Communications Benchmark, 150 human-labeled tasks, Surge AI, September 2026. Results are self-reported and an independent run is in progress.” Vendor claims on that set (product):

Claim	Figure	Attribution
Recall (“real violations caught”)	95.5% vs 92.9% best frontier tested	ZeroDrift self-reported
Speed / cost vs closest frontier models	Up to 34× faster, 12× cheaper	ZeroDrift self-reported
Latency	~1.5 s/message API; <100 ms self-hosted	ZeroDrift self-reported

GlobeNewswire repeats 95.5% and says Anchor outperformed GPT-5.6 Sol on recall, precision, and F1. SiliconANGLE notes attorney-labeled Surge AI data and that ZeroDrift published the benchmark, so the figures remain company claims (SiliconANGLE). Surge AI labeled the set; ZeroDrift scored and published — that is not third-party verification.

Specialized enforcement SLMs between agent outbound channels and FINRA/SEC-style communications rules — classify, cite, rewrite, log — at flat per-message prices. Lead on API availability and product shape; treat 95.5% / 92.9% / 34× / 12× as ZeroDrift’s own scorecard until an independent run lands.