LangChain Finds Jev 5–6x Faster Than Sonnet at Document Classification

An AI model need not run the whole job to make a workflow faster. In tests LangChain published September 25, TypeSafe AI’s Jev classified document pages faster than Sonnet and helped shorten a browser agent’s next-action delay. The design keeps open-ended work with a language model and sensitive calls with a person.

A model built for the branch, not the answer

TypeSafe introduced Jev on September 15 in early access. It takes information about a task and returns predefined, typed answers with probabilities rather than writing prose. That makes it a candidate for questions such as which route a document should take, not for a step that requires drafting an explanation.

LangChain’s new report puts those answers inside a larger workflow. In LangGraph, nodes perform work, state holds information from each step, and edges determine what runs next. Code can define the path while Jev supplies a judgment at a particular branch. The information accumulated in state also gives later steps their context; the model does not have to direct the entire job through a prompt.

That split also affects what happens when a run fails. LangChain says LangGraph saves state at each step, so work can resume with earlier decisions intact rather than start over and potentially take a different route. Its interrupts can pause a step for human approval, while LangSmith traces let a reviewer inspect the decisions made along the way.

Three questions for each page

LangChain illustrated the approach with a litigation document-review example. In one request per page, Jev assessed whether the page matched the document request, contained personal information, or might be legally privileged. A page that did not match was set aside. Personal information sent a page to a language model for redaction; possible privilege paused the workflow for an attorney. Pages left over were ready to hand over to the other side. LangChain says pages can be processed in parallel.

LangChain ran the same graph with Jev and with Sonnet handling classification. Across its trials, it says Jev was 5–6 times faster on that step. This is not a claim that the entire legal review finished 5–6 times sooner: redaction and attorney review remain separate work, and the reported comparison measures classification.

Choosing a browser action, then knowing when to hand off

The report’s browser example applies the same narrow-decision idea to Stagehand, Browserbase’s browser-automation tool. Stagehand marks interactive elements on a page, giving Jev a defined set of candidates rather than an open-ended instruction to operate the browser. Jev picks an action type and an element; a choice below a 0.7 confidence threshold falls back to a language model. In early testing, LangChain says median time for the act() operation fell from 1.97 seconds to 0.46 seconds, about 4.3 times faster.

The fallback is part of the design, not an exception to the result: Jev takes the bounded choices it is confident about, while the language model handles uncertain ones. The latency figure covers act(), the operation that chooses the next action. It does not describe how long an entire browser task takes.

A visible decision is not a verified one

LangSmith traces can show which route a document took and the probabilities behind Jev’s answers. Those controls make a decision inspectable; they do not, by themselves, show that it was right. The reported speed comparisons do not establish how often pages were routed correctly or whether the browser fallback caught mistaken choices. That distinction matters most where a fast classification determines whether a page is set aside, redacted, or sent to a lawyer.