AppZen grades its own finance models ahead of frontier AI on five of six expense-audit controls

Each system in AppZen's comparison got the same expense data, supporting documents and customer configuration. Each was scored on precision, recall and the combined F1 ^[7]. The evaluation is AppZen's own ^[7]. The announcement does not give the test-set size, the prompts, the model settings or an independent evaluation ^[8]. Of those four, the prompts matter most. ZenLM Plus is a family of finance-specific models ^[1]. A frontier model sees a customer's categories and thresholds only through a prompt someone wrote for it, and in this test AppZen wrote it. So the 9.1-point lead over GPT-5.6 Sol on policy-category cases ^[1] compares AppZen's model with AppZen's prompting of the seven frontier models ^[2].

For that lead to transfer, a buyer's expense data has to look like the test set. The models draw on fields employees enter, receipt extractions, merchant and card data, and each customer's configured categories and thresholds ^[12]. One case AppZen cites is separating a payment slip that shows a card was charged from a merchant receipt that substantiates what was bought ^[13]. A company whose staff mostly photograph card-terminal slips has a different document mix from one that submits itemized hotel folios. The 9.1-point win and the 0.8-point loss on receipt verification ^[3] come from the same test set of unstated size. Neither can be read with a margin of error.

The receipt-verification loss is the most credible result in the release. Gemini 3.1 Pro and Sonnet 5 each beat ZenLM Plus there ^[6], and AppZen designed for that outcome. Its Mastermind Platform routes work among its specialized models, deterministic rules and frontier models, then applies the output inside finance workflows and controls ^[14]. A router is the right design for a model that wins most controls and loses some. According to RuntimeWire, Kale, the chief executive, described that blended approach as part of the product, and Verma, the CTO, said the platform can use a frontier model for a task where it is the better fit ^[15].

The cost claim uses different comparison models from the accuracy claim. The strongest frontier accuracy result came from GPT-5.6 Sol ^[4]. The two cost comparisons AppZen gives are against GPT-5.6 Luna and Opus 5 ^[9]. By AppZen's own numbers, Opus 5 scored 85.2 on non-conforming receipt detection ^[2], against 92.4 for ZenLM Plus ^[5]. A team that routes receipt verification to a frontier model would be paying for Gemini 3.1 Pro or Sonnet 5, the two models that won that control ^[6].

In my view, the pilot that settles this runs one control at a time, on expense lines the team has already audited. The frontier comparison in that pilot should use a prompt the buyer's own staff wrote. RuntimeWire makes a similar point: customer deployments will test whether the claimed accuracy and cost advantages hold on real expense policies ^[16].