Product1 publisher3 min readPublished
ZeroDrift's own FINRA benchmark puts its 9B model level with GPT-6 Astra
The startup says Anchor 3.0 caught more than 90 percent of violations in a test it published itself, at under a 500th of the cost of a frontier call, and the messages it misses stay the firm's problem.
The Product Desk · Product desk
What happened
- ZeroDrift made Anchor 3.0 generally available through its Enforcement API, a family of small language models that inspect messages written by AI agents before those messages go out.
- The company reported that its flagship model detected more than 90 percent of violations in a benchmark built on Financial Industry Regulatory Authority rules.
- ZeroDrift said that model matched the overall accuracy of OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 while running more than 100 times faster at less than one 500th of the cost.
- The test data was attorney-labeled by Surge AI, which ZeroDrift says worked independently, but ZeroDrift published the benchmark itself, leaving the performance figures as company claims.
- The family ships in three versions: Mini and the flagship at 9 billion parameters with 4 billion active, plus a 27 billion-parameter Max for long documents, attachments and in-house policy.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- exposure A check that clears more than nine violations in ten still lets the remainder reach customers, and the firm's own supervisory procedure, not the vendor's benchmark, has to say what happens to those.
- constraint The cost advantage only shows up when the check runs on every outgoing message; a firm that samples can keep calling a frontier model and spend roughly the same money.
- contradiction Buyers cannot tell from the release which measure actually moved, since Mini is credited with catching more violations than both frontier models while the flagship is described only as matching their overall accuracy.
Something under one violation in ten got past the check, on the company's own headline number [1]. The person who signs the firm's supervisory procedures inherits that remainder. ZeroDrift was clear about this in its August preview, when it described the software as a risk-reduction layer and not a guarantee that every violation would be caught [13].
The cost claim is where the product case sits. Less than one 500th of a frontier call [3] is under 0.2 percent per check [2]. Read the other way, 500 Anchor checks cost less than a single call to GPT-6 Astra [3]. ZeroDrift did not publish a per-message price, so that is a ratio and not yet a budget line.
Kumesh Aroomoogan, ZeroDrift's founder and chief executive, said in a statement that frontier models have made it easy to build capable agents, and that the hard part is running them inside a regulated business, where every message has to follow the rules and the check has to happen every time, before anything goes out [4]. Agents can write thousands of customer communications faster than compliance staff can read them [14], and the every-time requirement is the condition the per-check price has to survive. The company recently added Guard for Agents, which puts the checks inside agent workflows through an API [12].
The other cost is false positives, which take reviewer time. ZeroDrift said Mini fired fewer than half as many false positives as Claude Fable 5.1 on its FINRA test [6]. A compliance lead can plan staffing around that number, since false positives set the size of the queue.
Two of the three tiers are the same size. Mini and the flagship are both 9 billion parameters with 4 billion active [7], so what the step up buys is rule coverage and editing [5]: more than 200 prebuilt rules spanning FINRA, the Securities and Exchange Commission and others, with the offending lines identified and rewritten before a message is released [8]. All three were post-trained from Google's Gemma E4B and Alibaba's Qwen3.8-27B [10].
Settle in writing, before a pilot, what share of outgoing traffic the check will cover; that decides whether a 500x cost gap shows up in your budget at all. Settle the miss rate you will accept, written as your own number and not as a vendor benchmark. And settle what happens when the model fires: the platform can flag, rewrite, block or route a message to a person, and it records the decision for later audit [11]. The rewrite path is the one that deserves an argument in the room, because it means a 9 billion-parameter model edited a customer message and sent it.
What to watch
- An independent run of the FINRA benchmark, or release of the labeled set, would test whether the accuracy parity with GPT-6 Astra and Claude Fable 5.1 holds outside ZeroDrift's own scoring.
- Per-message pricing for the Enforcement API would let a firm size the saving against its actual outgoing volume instead of a ratio.
- Whether buyers let Anchor rewrite and release customer messages, or restrict it to flag-and-route, once their audit teams read the logs.