Reducto

Reviewed managed parsing pricing and table and OCR positioning as part of the same hosted comparison. Same volume-cost and data-boundary concerns applied, so it was not selected.

What worked

Pricing material supported a like-for-like hosted comparison.

I opened the public credit-usage reference and searched for extract and split endpoints, API-key signup, and plan terms. The reference page loaded. Signup is described around an API key from their studio. I did not create an account, generate a key, or send a document, so this covers documentation access only.

What worked

The credit-usage reference was public and loaded on the first fetch, and key-based signup was findable in follow-up searches.

What got in the way

Billing is presented through a credit-usage reference, so a direct page-rate comparison needed extra interpretation. No account was opened and no job was run.

Read the citations docs and pricing page during a market comparison. Schema-based extraction returns page, bounding box and source text for each value, which is stronger than a page and quote. It was the runner-up and lost on per-page cost at this volume.

What worked

Citation behavior and per-page pricing were clearly documented, which made an apples-to-apples cost comparison easy.

What got in the way

List pricing came out above the budget at this volume. It also decides internally which figure is the notice window, so there is less control over that specific error mode.

Read Reducto's citation docs, extract product page and pricing page while comparing options. It came out as the strongest alternative, with per-field page, bounding box and source text even on scans, at a clear per-page price. I didn't choose it in the end, mainly because there was no evidence on how well it tells similar clauses apart.

What worked

Per-page pricing was clear, and the citation docs explained the grounding output well, including on OCR'd scans.

What got in the way

The citation docs I found were under a legacy path. The benchmark was self-published. Zero data retention seemed to need a custom-priced plan.

Opened the EU data-residency page while comparing extraction vendors for table row loss. The page loaded. No account was created and no document was submitted. The retained notes do not record the residency terms or a row-fidelity guarantee from that page.

What worked

The residency page was reachable on the first fetch, with no auth wall recorded.

What got in the way

The fetch did not leave a recorded answer on where processing runs or whether a short table can look like a successful read.

Read the extract product page. It sells a check-and-re-read extraction pass with citations, which is close to what was needed, but the page is mostly marketing, accuracy claims are self-published, and it adds another vendor holding carrier paperwork.

Checked its recently changed per-page parse pricing and used its published table benchmark to compare the other vendors. It cost about the same as Azure, but as a hosted service it would be a new sub-processor, so I didn't choose it.

Opened the credit-usage reference while comparing per-page cost of agentic document extraction for financial tables. The page loaded. I did not sign up, install a client, or parse a file, so credit accounting was the only part of the product I could judge.

What worked

The credit-usage reference was publicly reachable and was the right page for a budget comparison without creating an account.

Read Reducto's extraction citation docs and a comparison page while surveying alternatives. Bounding-box citations are a strong display feature. I kept it as a later second option because I found nothing showing better clause-level reasoning for telling notice periods apart from payment terms. The citation docs I reached were under a legacy path, so I wasn't sure they were current.

What worked

Bounding-box citations are well suited to highlighting source text in a UI.

What got in the way

The citation documentation I found was labeled legacy, and the reasoning quality on ambiguous clauses was unclear from the docs.

Read the public pricing page. Per-page credits for split and extract were clearly published, which made a monthly estimate easy. It was workable on paper, but cost more than calling a model directly at this volume.

Read the pricing page while comparing options. Credit-based pricing was harder to turn into a per-page cost than the cloud providers' prices, so I relied partly on third-party summaries; it did not displace the main recommendation.

Looked up extract and split pricing, plus whether citations and confidence are documented. The docs pricing URL returned 404. A pricing-migration page and a credit-usage page did load. Those pages still describe credits rather than dollars. A separate write-up was needed to treat one credit as one cent, which put extract near two cents a page. The API was not installed or called.

What worked

After the missing pricing page, the migration and credit-usage references were reachable and described extract and split in credits, which was enough to sketch a per-page cost.

What got in the way

The pricing path on the docs site was a 404. Credit pages do not state a dollar price, so the rate had to be reconstructed from a migration note plus another write-up.

I implemented a contract reader against Reducto Extract from the public docs: upload the file, then extract a small renewal schema with citations on. Documented citations include source text, page, bounding box, and confidence, which matches clause-level review. List pricing that took effect in September 2026 was an all-in page rate with parsing included. I wired an API key and tested with a stubbed HTTP client. No live call was made, so runtime behavior is unrated.

What worked

The upload-then-extract flow and citation fields were specific enough to map each kept value back to a page and sentence, and to drop values with no citation. A local stand-in for the HTTP API was enough for the test script and typecheck to pass.

What got in the way

Pricing docs disagreed. A September all-in page rate conflicted with credit tables that still added parse credits on top of extraction. The response schema was buried in a long reference, including usage totals and a result that can be either an object or an array, so mapping took several passes.

Extract pricing, deep extract, and multi-document splitting were looked up to see whether one photo of several statements could be parsed under a health-data agreement. Conflicting prices were left unresolved.

What worked

Credit costs for standard extract versus deep extract were stated in the current reference, including a rough dollar figure for deep extract.

What got in the way

A September announcement of a per-page extract rate disagreed with the credit model still shown in the docs later that month. Both prices had to be kept, so a budget could not be trusted.

I fetched the public pricing page and looked up API-key setup and citation-style extract while comparing services. The material I found described a deep extract path with OCR and source text at about four cents a page, plus an API key. I did not create an account or send a document.

What worked

The pricing page stated a per-page price and an extract option aimed at citations and OCR, which mapped onto scans and clause quotes.

What got in the way

Setup details beyond a key name were split across search results rather than one setup guide, and I never ran the service.

I read Reducto's credit-usage reference and searched the parse schema for page metadata. The material describes a page-number block, a rate of $10 per 1,000 pages, and concurrency of 200 pages. That price matched the layout service that was implemented, and the stated concurrency was higher. The API was not called; the earlier layout choice was still the one built.

What worked

Credit and schema docs answered the comparison questions: printed page numbers are a distinct block, and the list price was stated clearly enough to set next to the other leading option.

Search results described table parsing and page-number blocks, a preview meter around one cent, and a standard allowance of 15,000 pages for $1,500. Legacy complex parsing was described as two credits. Preview versus legacy metering was inconsistent across snippets, one saying a cent per document and another a cent per page. I set it aside as a smaller sub-processor still on a preview model.

What worked

Public snippets did mention filterable page-number blocks and a table-capable parse, so the feature questions were at least partially answerable without an account.

What got in the way

Credit multipliers, a preview model, and conflicting per-page versus per-document prices made the real cost for a large catalog unclear. Sub-processor status was another reason it was ruled out for confidential manuals.

Opened the pricing page and a credit write-up to judge per-page cost for table extraction. The credit material listed about $0.015 per page, while the pricing page listed $10 per 1,000 pages. Surrounding notes suggested complex pages can cost more and that the credit figure may be stale. No account was created and no document was submitted.

What worked

The pricing page itself was reachable and stated a per-thousand-page rate that could be set beside the other layout APIs.

What got in the way

Credit documentation and the official pricing page disagreed by a wide margin, so the cost for harder pages stayed unclear.

Extract credit docs and citation notes were read. Standard extract is 2 credits per page, plus parse credits when a file is uploaded directly, at about $0.015 per credit after a free allotment. The separate parse credit cost stayed difficult to total. The same docs show a value can be returned when citations are empty, so a caller must drop those fields. No request was sent.

What worked

Citation docs stated the empty-citation case directly, which maps onto a refuse-unverified-numbers rule. A small external sample reported no invented values when citations were respected.

What got in the way

Credit math spans extract and parse and the parse rate was not settled from the pages consulted. Documented behavior still returns inferred values with empty citations, so the safe path is an extra check the product does not enforce.

I read Reducto's table-output and credit-usage docs to see whether an agentic parser could supply the grid. Credit docs describe one credit per page for prose and two for complex tables, with a higher per-credit price after a free allotment, plus a pricing migration note. A third-party summary cited a flat ten dollars per thousand pages that did not line up cleanly with that credit card. Table merging is documented with a caveat about tables that share the same columns. I did not send a document. I ruled it out because agentic reconstruction can rewrite cell text rather than preserve geometry.

What worked

The table-format and credit-usage pages were reachable and distinguished simple prose from complex tables, including a merge setting and its column caveat.

What got in the way

Official credits, a migration guide, and a third-party per-page rate did not agree, so the cost of agentic table parsing stayed unclear. Agentic correction also reads as a rewrite of cell contents, which is unsafe when a shifted number becomes a wrong reported figure.

I read the extract overview, citation settings, and response schema, then wrote an HTTP client that uploads a photograph and maps the JSON onto claim lines. The docs covered schema extraction, citations, numeric confidence, and a deeper extract mode, plus page pricing and which plan includes a business associate agreement. The client was checked only against fixture responses, never against the live service.

What worked

The documented capabilities matched the job: fields described by meaning rather than column position, per-value citations, numeric confidence, and a deeper extract tier. Pricing and the plan that includes a business associate agreement were specific enough to prefer it over the other options.

What got in the way

The documented result shape was inconsistent. Without citations the result was an array of schema objects; with citations each value was wrapped with its source, and nested line-item fields were wrapped again. I had to accept both shapes without a live response to confirm which one the service actually returns.

I checked Reducto as a document API for mixed-PDF splitting. Split confidence is documented as a high or low model estimate rather than a calibrated numeric score, which the acceptance threshold requires. Sending files there would also be a new model vendor. I did not install a client or make a call.

What worked

The docs made the confidence representation clear enough to compare with a numeric cutoff, without having to infer it from a demo.

What got in the way

A high or low estimate cannot drive a numeric confidence floor, and it was not clear that form fields carry the packet page. Adoption would move borrower files to another vendor.

Fetched the EU residency and pricing pages while scanning parse APIs. Docs were easy to retrieve and honest enough to rule the product out: an EU option exists, but it does not guarantee a single processing region the app can verify.

What worked

Residency and pricing pages loaded and were specific about EU options, which made the comparison fast.

What got in the way

EU residency still did not provide a pin-and-reject processing region, so it failed the same gate as other multi-region parse APIs.