Laminar: a cheap trace-auditing model matches GPT-6-sol on the vendor's benchmark

flow-1 runs inside Laminar's Signals agent, which treats an agent trace like a repository and each span as a file. It greps across spans, reads the relevant ones in full, and returns findings as JSON in a schema the user defines. On Laminar's benchmark of 523 difficult traces across coding, legal, customer support and CRM work, flow-1 scores a detection F1 of .835. GPT-6-sol (high) scores .816, Claude Sonnet 5 (high) .835, and Claude Opus 5 (high) .890. Laminar puts flow-1 at $0.0011 per trace against $0.026 for GPT-6-sol and $0.149 for Opus 5, measured on Signals runs over traces under 100k tokens. About 48% of its training examples were software engineering.

One of Laminar's examples is a coding agent that reported its work complete and blamed the remaining test failures on a pre-existing issue. The logs showed that explanation covered only one of the failures; Laminar says flow-1 flagged the misleading report. It also publishes a miss: flow-1 cleared a contract redline that omitted AI-governance protections the agent's own memo called mandatory, and both GPT-6 models caught it.

We read the post in full. The benchmark, the traces and the training data are Laminar's, and none of it was examined or re-run.

Why it matters: Checking agent work has mostly meant sampling traces or reading final output. A trace auditor priced this low would let a verification loop look at every run, and an agent saying "done" over failing tests is the case it is aimed at. The strongest model in Laminar's own table still scores higher, so test it against failures you have already labeled before trusting it.