Benchmarks

Every score, document by document

Per-page and per-table scores for all three systems are in results/, and the scripts that ran them in research/benchmarks/.

Three systems on the same documents with the same requests (November 2026):

• Reducto r-1: Reducto's hosted API, default settings.
• Model only (Haiku 5.5): ReadAnyDoc's model-only profile: one Claude Haiku 5.5 read per page, nothing else.
• Engine (Haiku 5.5): ReadAnyDoc's engine profile, the default. A test in tests/test_pipeline.py checks that the shipped profiles are exactly the configurations measured here.

Every score comes from the benchmark's own grader or a deterministic script; no model judges the output. 95% confidence intervals are bootstrap intervals over documents; "paired" differences resample the same documents for both systems. Pages used while developing ReadAnyDoc are reported separately as "the sample"; the headline numbers hold on the pages never seen during development.

Results

Benchmark	Measure	Reducto r-1	Model only	Engine
olmOCR-bench, 1,403 pages	official score, %	80.9	82.2	84.2
	multi-column	65.6	79.8	81.0
	long tiny text	84.2	93.2	93.9
	headers and footers	90.1	70.4	81.2
	old scans	49.0	47.5	50.4
	tables	92.7	91.7	92.2
OmniDocBench, 1,651 pages	text edit distance (lower is better)	0.095	0.081	0.081
	table TEDS, %	87.9	85.4	87.0
	reading-order edit distance (lower)	0.161	0.167	0.170
	formula edit distance (lower)	0.133	0.158	0.156
	layout: block type right, %	84.6	55.6	78.3
	layout: block found at IoU >= 0.5, %	67.2	26.5	61.1
RD-TableBench, 1,000 tables	Reducto's own table score	88.85	88.20	88.17
FUNSD, 199 forms	question/answer links found, %	78.6	71.6	77.4
LongExtract, 10 long documents	table rows extracted, %	53.9	75.4	75.9

What the differences mean, using the paired intervals:

• The engine against one model read: clearly better on layout (+23 points block types, +35 boxes found), forms (+5.8), headers and footers (+11) and the olmOCR overall score (+2.0); level on tables, plain text and long extraction. The engine costs about 2-3x as much per page.
• The engine against Reducto: ahead on olmOCR (+3.3; multi-column and small print), on OmniDocBench text and on long extraction; level on RD-TableBench and FUNSD (differences inside the intervals); behind on layout (-6), formulas, header and footer removal, and speed.

Cost and speed

	Reducto r-1	Model only	Engine
Cost per page	$0.010 (list price)	$0.001-0.002	$0.003-0.005
Seconds per page (median, tables)	4	8	20-50

ReadAnyDoc costs are Anthropic API charges measured on these runs. Scanned pages with tables and handwritten pages cost more (up to about $0.03 a page on handwriting, which goes to Sonnet 5.5).

Method and caveats

• Same input files and same request to every system. Reducto was called through its public API in November 2026.
• olmOCR-bench is scored with its official benchmark.py; RD-TableBench with Reducto's published grading code; OmniDocBench and FUNSD with scripts that apply the benchmark's definitions to each system's blocks; LongExtract counts expected table rows found in the extraction result.
• LongExtract ran on the 10 shortest documents of the set for all three systems; Reducto alone ran on 32.
• Between 1.5% and 3% of pages in these sets came back empty from Claude because the provider declined to return their text (content filtering). They are scored as empty, which lowers both ReadAnyDoc columns; the response reports them in warnings.
• Only Claude Haiku 5.5 (with Sonnet 5.5 for handwriting) was measured. Other models work with ReadAnyDoc; their accuracy is unknown until measured.
• Other parsers (Mistral OCR, LlamaParse and so on) were not run.
• Benchmark data is not redistributed here; each benchmark's own license applies. The scripts that ran these numbers are in research/ (see research/README.md).