ZeroDrift has launched Anchor 3.0, a family of small language models that review AI agent messages before they are sent. Its flagship model detected more than 90% of violations in a benchmark based on Financial Industry Regulatory Authority rules. The models are available through ZeroDrift's Enforcement application programming interface. The move matters because autonomous agents can generate thousands of customer communications faster than human compliance teams can review.
Anchor 3.0 lineup and test results
ZeroDrift said the flagship model matched the overall accuracy of OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 in its FINRA test. At the same time, it ran more than 100 times faster and at less than one 500th the cost of those frontier systems. The test used attorney-labeled data produced independently by Surge AI, according to the company. ZeroDrift published the benchmark itself, so the results remain company claims. Founder and Chief Executive Kumesh Aroomoogan said capable agents became easy to build with frontier models, while the hard part is operating them inside a regulated business with a check before each message goes out.
Anchor works as a pre-send layer that intercepts communications and checks them against regulations and company policies, with options to flag, rewrite, block or route messages for human review while logging decisions for audit. The lineup includes three versions. Anchor 3.0 Mini is a 9 billion-parameter mixture-of-experts model with 4 billion active parameters for high-volume traffic and prebuilt rule packs. The flagship Anchor 3.0 uses the same parameter count but applies more than 200 prebuilt rules covering FINRA, the Securities and Exchange Commission and other regulations, identifying violating lines and rewriting them. Anchor 3.0 Max is a 27 billion-parameter model for long documents, attachments and company-specific policies without additional fine-tuning, and the models were post-trained from Google's Gemma E4B and Alibaba's Qwen3.8-27B.
The release follows an August preview in which ZeroDrift said Anchor combined deterministic checks with an open-source model trained on regulatory material and attorney-labeled communications. At that time the company described the software as a risk-reduction layer rather than a guarantee that every violation would be caught. That framing fits the current problem: frontier systems lowered the cost of building capable agents, but regulated firms still need a check on every outgoing message. ZeroDrift also recently introduced Guard for Agents, an API-based service that embeds compliance checks directly into agent workflows instead of leaving review to separate manual steps.
What this means for regulated businesses
For businesses, the practical effect is coverage of every message without adding manual review capacity. A compliance check that runs in real time and at low cost makes it feasible to screen high-volume agent traffic, including routine customer replies that previously went out unchecked. Small firms without large compliance teams can start from prebuilt rule packs for FINRA and other requirements, while larger financial firms can apply more than 200 rules and keep an audit record of each decision. The difference shows up in daily operations: fewer bottlenecks before sending, plus a consistent record of screening outcomes.
The limits center on verification and fit. Performance figures come from a ZeroDrift-published benchmark, even though labeling came from Surge AI, so buyers need independent testing on their own message types and policies. False-positive behavior matters as much as detection rates, since excessive flags create extra review work and slow down response times. Coverage also differs by version, from high-volume packs in Mini to long documents and company-specific policies in Max without additional fine-tuning. Before choosing, companies should clarify latency under peak load, how rewrites are logged, how Guard for Agents and the Enforcement interface connect to existing workflows, and what happens when a rule has no clear match.
A useful marker will be production evidence from regulated deployments: independent FINRA-related accuracy numbers, false-positive rates and audit results over several months. If financial firms keep the system in front of live agent traffic and expand from prebuilt packs to company policies, the approach will have moved from benchmark claims to operating practice.
