Differences and limits

What is different from Reducto, so nothing surprises you after switching.

Behaviour

Area	Reducto	ReadAnyDoc
Speed	about 5 s a page	20 to 50 s a page with the engine, 8 to 28 s model only; pages of one document are read in parallel
Page layout	ahead by about 6 points on OmniDocBench block types	78.3 against 84.6
Formulas	ahead	behind (formula edit distance 0.156 against 0.133)
Headers and footers left out	90.1	81.2
Text, multi-column pages		ahead (olmOCR-bench 84.2 against 80.9)
Pages the provider will not return	not applicable	1.5% to 3% of pages in large public sets come back empty because the model provider declines them (content filtering). The response says so in warnings
Confidence		tiers from how a value was found, not calibrated probabilities

Response fields

• usage.credits is always null (no billing); studio_link is null.
• pdf_url is set for PDF inputs only.
• usage.model_usage (tokens per model) is an addition.
• Extract responses always include warnings.

Not available

• POST /pipeline and /pipeline_async (deprecated by Reducto) return 501 with advice to chain /parse and /extract.
• /configure_webhook has no Svix portal; it returns and sets the default webhook URL.
• settings.timeout, tenant_throttling, force_file_extension, hybrid_vpc are accepted and ignored.
• No hosted Studio. Use the library, the CLI or your own UI.

Known issues

• A small number of table-heavy documents can take much longer than average; READANYDOC_DOCUMENT_TIMEOUT_SECONDS (30 minutes by default) bounds it.
• On very long list extractions, one run in our tests lost most rows; array_extract=True and checking warnings are recommended for long tables.