Compare
Reducto vs LlamaExtract: price, confidence scores, grounding, benchmarks and terms, with Velrim in the same table
Reducto and LlamaExtract are the two extraction platforms most often shortlisted together, and both have terms that prohibit the use a benchmark run would require. This page puts them side by side on what each publishes, with Velrim in a third column so the comparison includes the vendor writing it. Every Reducto and LlamaExtract entry is a quotation from that vendor's page with the day it was read.
Side by side
| Reducto | LlamaExtract | Velrim | |
|---|---|---|---|
| What it is | Parse, Split, Extract, Classify, Edit and Studio. "The agentic document platform". | Extract on LlamaCloud, beside Parse, Classify and Split. "Extract structured data from documents using custom schemas." | One extraction API. Schema in, typed fields out, each with a page location and a 0 to 1 confidence score. |
| Extraction list price per 1,000 pages | Extract $20. Deep Extract $40. Standard tier. | Cost-effective $10.00. Agentic $31.25. Agentic Plus $75.00. Turbo $43.75. Credits at $1.25 per 1,000, extract plus parse. | $20, grounding and confidence included. |
| Free tier | $150 in free usage on Standard. | 10K credits: 400 Agentic pages. | ~500 pages, no card, 90 days. |
| How to buy | Standard self-serve. Growth and Enterprise: "Custom pricing", "Contact sales". | Free, Starter $50 a month, Pro $500 a month, Enterprise "Custom". | Self-serve. Prepaid balance from $5. |
| Per-field confidence | With citations on: confidence as "high" or "low", plus extract_confidence and parse_confidence as 0 to 1 numbers. | With confidence_scores: true: parsing_confidence, extraction_confidence and a combined confidence. | One 0 to 1 score per field, always on. |
| What the docs say the score means | "Low parse_confidence suggests OCR errors. Low extract_confidence suggests the model was uncertain about interpretation." | "a score approximates a real probability of correctness" on three tiers. "At a 0.8 threshold roughly 75% of extraction errors fall below the line." | A probability that the field is correct. Claimed against observed differed by 13 points on average in the published comparison. |
| Evidence for the score | None published on the citations, extract overview or response format pages as of 2026-09-03. | None published on the extensions, configuring or product pages as of 2026-09-04. | Error curves, regenerated on every refit. Per-bucket table and raw outputs in the published comparison. |
| Grounding | "Return page number, bounding box, and source text for each extracted value." Citations cannot be combined with chunking. | Page number, verbatim text, box x, y, w, h and page dimensions with cite_sources: true. On Turbo, text-level only. | Page, bounding box and matched snippet on every field, from a word map, no second model call. |
| Benchmarks published | RD-TableBench, 1,000 table images, dataset and grading code public (Nov 2024). Ranked first on the micro1 long-extraction benchmark (Jun 2026). | ParseBench, ~2,000 pages and 167,000 test rules, 14 methods, LlamaParse Agentic 84.9% (Apr 2026). Reducto is one of the 14. | A pre-registered comparison of six extraction setups on 124 public documents (Sep 2026). Raw outputs and scorer public, DOI 10.5281/zenodo.22233430. |
| Terms on benchmarking | "Access or use the Services for any competitive purpose." is a prohibited use. Effective 17 April 2026. | "for benchmarking or competitive analysis of the Service" is a prohibited use, §2.2(h). Last updated Jun 7, 2024. | None. The eval CLI ships adapters for both. |
| SDKs | Python and Node, CLI, MCP server. | Python, TypeScript, Go, Java, CLI, REST, web UI. | TypeScript and Python, MCP server. |
| Funding | $108M total, $75M Series B led by a16z (Oct 2025). | $27.5M total, $19M Series A led by Norwest (Mar 2025). | Self-funded. |
Confidence scores, in each vendor's words
Each citation includes aconfidencefield with a categorical value ("high"or"low").
By default in v3,numerical_confidenceis enabled, which addsgranular_confidencewith numeric 0-1 scores.extract_confidence: How confident the LLM is about the extraction.parse_confidence: How confident the OCR/parsing is about the underlying text.
Lowparse_confidencesuggests OCR errors. Lowextract_confidencesuggests the model was uncertain about interpretation.
This feature adds three confidence-related fields to the extraction metadata:parsing_confidence: Confidence score indicating how well the relevant context was parsed from the source document.extraction_confidence: Confidence score indicating the relevance of the extraction based on the JSON schema field.confidence: Combined confidence score that incorporates both parsing and extraction confidence.
Calibrated on Cost Effective, Agentic, and Agentic Plus. On those tiers a score approximates a real probability of correctness, so you can set a threshold directly rather than only ranking fields against each other. At a 0.8 threshold roughly 75% of extraction errors fall below the line.
Agentic Max and Turbo return scores from an earlier model. They are still useful for ranking fields, but the calibration above does not apply to them.
Reducto returns two 0 to 1 numbers and a high or low label, and says what a low value suggests. LlamaExtract returns three numbers and says the combined one approximates a probability of correctness on three of its five tiers, with one figure attached, 75% of errors under 0.8. As of 2026-09-04, neither vendor's pages carry a dataset, a sample size or an error curve for its score. One number per field, 0 to 1. In Velrim's published comparison of six extraction setups (September 2026), across every field the score graded, the claimed confidence and the observed share correct differed by 13 points on average with refusals counted as a coin flip, 10 points without them. On the 33 values Velrim itself invented, the score averaged 0.40 against 0.64 overall. The curves are on the error curves page and regenerate on every refit. The definition of the fabrication measure is on the metric page.
Grounding
Return page number, bounding box, and source text for each extracted value. Useful for verification and debugging.
The page number where the information was found. The verbatim text that was used to extract the field value. Bounding box coordinates (x,y,w,h) indicating the exact location of the cited text on the page. Page dimensions (width,height) to help you render the bounding boxes accurately.
Cite sources: Enable cite sources to attach citations to extracted fields so you can trace every value back to its origin in the document. On Turbo, citations are text-level and carry no bounding boxes.
Both return a page number, a box and the source text per field when citations are on, and both document a cost for turning them on: Reducto's citations cannot be combined with chunking, and LlamaExtract's docs say citations and confidence scores "will significantly slow down extraction processing time". Velrim returns page, box and matched snippet on every field from the PDF's own word map, with no option to turn off and no second model call.
Benchmarks each vendor publishes
1000 complex table images from a diverse set of publicly available documents, manually annotated by a team of PhD-level human labelers. Dataset at huggingface.co/datasets/reducto/rd-tablebench, grading code at github.com/reductoai/rd-tablebench. November 4, 2024.
micro1 released an independent benchmark evaluating document extraction systems on difficult, high-field-count workloads. Reducto Deep Extract ranked first overall, 100% completeness, 99.6% precision, 99.6% recall, and 99.3% leaf accuracy. June 30, 2026.
Today we're releasing ParseBench, a benchmark of ~2,000 human-verified enterprise document pages with over 167,000 test rules, organized around five capability dimensions: tables, charts, content faithfulness, semantic formatting, and visual grounding. We evaluated 14 methods spanning vision-language models, specialized document parsers, and LlamaParse. While no single method is good at everything, LlamaParse Agentic was the only method tested that is competitive across five key dimensions, scoring 84.9% overall. Apr 13, 2026.
We tested 14 methods across three categories: general-purpose VLMs (GPT-5 Mini, Haiku 4.5, Gemini 3 Flash, Qwen 3 VL, Dots OCR 1.5), specialized document parsers (Textract, Azure Document Intelligence, Google Cloud Document AI, Reducto, Docling, Extend, LandingAI), and LlamaParse (Cost Effective and Agentic modes).
Reducto's benchmark scores tables and LlamaIndex's scores parsing, and each vendor's own product leads its own table. ParseBench lists Reducto among the 14 methods it scores. Neither is an extraction benchmark with a public answer key for absent fields. Velrim's published comparison of six extraction setups (September 2026) is that, on 124 documents with 96 audited absent fields, and neither vendor is in it, for the reason below.
Terms
Access or use the Services for any competitive purpose.
Except to the extent a restriction is prohibited by Applicable Law, you will not do, and will not assist, permit, or enable any third party to do, any of the following:
(h) use or display the Service in competition with us, to develop competing products or services, for benchmarking or competitive analysis of the Service, or otherwise to our detriment or disadvantage;
Reducto's terms are dated "Effective: 17 April 2026" (Reducto Terms of Service, read 2026-09-03). LlamaIndex's carry "Last Updated: Jun 7, 2024" (LlamaIndex Terms of Service, read 2026-09-04). Both clauses are use-based, so a paid account does not cure them, and the LlamaIndex preamble extends the ban to third parties. Velrim sent both vendors a written consent request on 2026-07-12 with a deadline of 2026-07-20. Neither replied by the deadline, and the pre-registered outcome for silence was to drop the arm and disclose. The record is in the analysis plan. Both adapters remain published in velrim-eval for a buyer running them under their own account.
What the comparison measured
| Arm | macro-F1 (norm) | fabrication (absent fields) | wrong among the top 90% by confidence | $/1k pages |
|---|---|---|---|---|
| Velrim | 0.691 | 11.5% [6.1, 19.8] | 36.1% [29.9, 44.1] | $20 (list; measured matches) |
| Gemini free-decode | 0.709 | 17.0% [10.1, 26.7] | not requested | ~$2.2 (measured, token-priced) |
| Gemini constrained | 0.697 | 12.8% [7.6, 20.5] | not requested | ~$2.2 (measured, token-priced) |
| gpt-5.4-mini free | 0.735 | 10.8% [5.7, 18.4] | none surfaced | ~$1.0 (measured, prompt-cached) |
| gpt-5.4-mini structured | 0.718 | 10.4% [5.3, 17.6] | none surfaced | ~$1.0 (measured, prompt-cached) |
| Mistral OCR 4 | 0.712 | 40.3% [31.0, 52.5] | none surfaced | $5 (list; measured $5.3) |
Six setups on 124 public documents, 96 of their fields absent by the audited answer key. Pooled, the six invented a value for 17% of those fields. Velrim: 11.5% [6.1, 19.8]. Velrim is the most expensive row and ties on accuracy. The full comparison has the intervals, the per-document-type breakdown and every raw output.
Pricing
| Product | List price per 1,000 pages | Basis |
|---|---|---|
| Reducto Extract | $20 | Standard tier, per page |
| Reducto Deep Extract | $40 | Standard tier, per page |
| Reducto r-1 Parse | $10 | Standard tier, per page |
| LlamaExtract Cost-effective | $10.00 | 8 credits a page at $1.25 per 1,000 credits |
| LlamaExtract Agentic | $31.25 | 25 credits a page |
| LlamaExtract Agentic Plus | $75.00 | 60 credits a page |
| LlamaExtract Turbo | $43.75 | 35 credits a page, no parse step |
| Velrim | $20 | $0.02 per extracted page, grounding and confidence included |
Standard: Free, $150 in free usage. r-1 Parse $10, Extract $20, Deep Extract $40, Split $20, Deep Split $40, Classify $7.50, Edit $60 or $15 for pre-filled pages, per 1,000 pages. Growth and Enterprise: Custom pricing, Contact sales.
Extract tier, Extract credits/page, Default parse tier, Parse credits/page, Default total: Agentic Plus, 50, Agentic, 10, 60. Agentic, 15, Agentic, 10, 25. Cost-effective, 5, Cost-effective, 3, 8. Turbo, 35.
Free $0 /month, Includes 10K credits. Starter $50 /month, 40K credits. Pro $500 /month, 400k credits. Enterprise, Custom. 1,000 credits = $1.25.
Reducto prices each endpoint per page on its Standard tier with $150 in free usage. LlamaExtract prices in credits, extract tier plus parse tier, so Agentic is 25 credits or $31.25 per 1,000 pages, and the free plan's 10K credits are 400 Agentic pages. Velrim is $20 per 1,000 extracted pages, prepaid from $5.
Which to pick
Reducto, for parsing, splitting, classifying and editing at enterprise volume, with a sales team and a Studio for review. LlamaExtract, for a schema builder in a web UI, a choice of tiers, and a free plan with no card on the same platform as LlamaParse. Velrim, when the requirement is a number per field to sort on and the evidence for that number in the open, at a list price with no sales call.