Datalab ranks 95.9% on a 225-document extraction benchmark at under half Reducto's price

• According to VikParuchuri's benchmark post, Datalab's Balanced extraction mode scored 95.9% on the company's internal benchmark, versus 95.1% for Reducto Deep Extract.
• The launch post also claimed the higher score came at less than half Reducto's price, but both the accuracy and price comparison are vendor-reported.
• In a thread reply, VikParuchuri's benchmark-details post said the benchmark covers 225 documents, including long documents, complex schemas, and failure cases that have tripped extraction models before.
• The main post said the updated mode now returns full verification with citations and reasoning, aimed at showing which extracted fields still need manual review.
• VikParuchuri's access post said users can try Balanced mode in the Extract playground, with $20 in free credits for signups using a work email.

You can open the Extract playground from VikParuchuri's access post, inspect the benchmark framing in the launch thread, and note that the only methodology detail surfaced so far is the 225-document description, not a public per-document breakdown.

The headline number is narrow but concrete: Datalab says Balanced mode reached 95.9% on its internal extraction benchmark, while Reducto Deep Extract scored 95.1% on the same test set, according to VikParuchuri's benchmark post. The same post tied that accuracy claim to a price claim, saying Balanced mode costs less than half as much.

That combination, slightly better extraction and a much cheaper list price, is the whole story here. The catch is also straightforward: the comparison comes from the vendor that shipped the update, not from an external benchmark run.

The more interesting product detail is not the 0.8 point edge over Reducto. In the launch post, VikParuchuri said Balanced mode now includes full verification, with citations and reasoning attached to extracted values so users can see which fields deserve manual inspection.

That makes the feature sound closer to a reviewable extraction workflow than a single opaque score. The evidence provided here is still just the announcement and attached demo video, with no separate documentation link for the verification format.

A follow-up post from VikParuchuri added the only public methodology detail in the thread: the benchmark uses 225 documents, includes long documents, and is built around complex schemas plus specific error cases the team has seen confuse extraction systems.

That does not answer the usual benchmark questions, like document mix, scoring rubric, or whether the Reducto run was reproduced under identical settings. It does at least establish that the claim was not framed around a tiny toy set.

VikParuchuri's access post said the feature is live in Datalab's Extract playground and comes with $20 in free credits for users who sign up with a work email. The thread positions Balanced mode as an immediately testable product update, not a preview or waitlist.