Introducing r-1 mini: premium parsing without the premium price

Today, we’re launching r-1 mini: high-quality document parsing with better performance than alternatives 25x the cost.

Reducto began by making the hardest documents usable by AI. With r-1 mini, we’re working toward a future where all document data can be put to use.

r-1 mini breaks a new price barrier for high-accuracy document parsing at $1 per 1,000 pages, outperforming both cost-effective and premium alternatives up to 25x the cost.

In our evaluations, r-1 mini outperformed leading cost-effective and premium document parsers: the next higher-scoring dedicated parser cost 10x as much, while the next-cheapest full-feature alternative cost 3.75x as much and scored 16 percentage points lower on accuracy

R-1 mini offers agentic capabilities for complex docs, at prices others alternatives only offer raw OCR. Read the full results and more below.

Why did we build r-1 mini?

As AI applications and agents take on longer, more complex tasks, they need accurate, structured information to work from. High-quality document parsing makes that information accessible, but its cost can make otherwise valuable use cases prohibitively expensive at scale.

We recently introduced r-1, our unified model that delivers full agentic parsing capabilities out of the box for $10 per 1,000 pages. R-1 was already a major step change for customers, who saw most major edge cases solved at a fraction of the cost of agentic parsing products. However, we saw amongst our enterprise customers that with the volume of context they needed to process, lower priority documents still had to be skipped or ignored due to cost constraints.

We built r-1 mini so every AI application can start with the best possible data: accurate enough to trust and affordable enough to use on every page. Our unified architecture gave us the foundation to train a smaller model that preserves strong parsing quality at a much lower cost. 

Reducto r-1 mini costs $1 per 1,000 pages, at a higher quality than any other cost-effective alternative. That gives teams room to:

• Make entire unstructured data archives processable at a higher quality.
• Add document sources and context without a proportional increase in spending.
• Consolidate and simplify their processing pipelines, at a lower cost.
• Build applications whose document volumes previously made them too expensive to serve (web crawling, research agents, enterprise knowledge search, etc).

How r-1 mini compares with other document parsers

Introducing ParseArena and PAIR

We built ParseArena to compare document parsers on a diverse enterprise document set, evaluated by PAIR (Pairwise AI Ranking), our open-source AI judge framework.

We’ve seen other document benchmarks rely on rigid, deterministic scoring rules that can penalize valid outputs simply because they differ from an expected format. We found that even small changes to existing document benchmarks, such as slight rotation or added padding, can substantially affect performance – suggesting that some models may have been trained specifically on those benchmark datasets.

Enterprise documents have a long tail of layouts, structures, and nuances that those rules struggle to fairly evaluate.

PAIR compares two parser outputs against the original page, evaluating which more faithfully preserves its information. This approach accommodates variations without requiring a hand-written scoring rule for every edge case. PAIR also enables a more holistic evaluation of a parse as a whole, as opposed to components in isolation.

We encourage every company to evaluate parsers on its own documents. Open-sourcing PAIR makes it easier to run private evaluations without publishing sensitive documents or results. Our team of experts is equipped to support your internal evaluations step-by-step, within the privacy of your own environment.

We’ll share a full paper and dedicated blog post on ParseArena and PAIR soon, covering the methodology, detailed results, and how to run your own evaluations out of the box.

ParseArena Results

ParseArena evaluates document parsers on 2,000 public-document pages spanning seven industries and six page types.The head-to-head evaluation compares outputs with vendor names hidden and allows ties.

The results make a strong case for reconsidering your default parser:

• Reducto r-1 mini outperformed Azure Layout and Amazon Textract Layout + Tables overall in head-to-head comparisons, at a fraction of the price. 
• • It also won more comparisons than it lost against Gemini Layout Parser, the closest cloud competitor in the evaluation, at a 90% cheaper rate.
  • r-1 mini produced 60% fewer serious errors per page than Amazon Textract (Layout + Tables), and about half as many as Azure Document Intelligence.
• Reducto r-1 mini delivered stronger overall head-to-head results than Mistral OCR 4.1, LlamaParse Cost-effective, Datalab Fast, and Extend Light, while costing less than each tested configuration.
• Reducto r-1 mini also outperformed LlamaParse Agentic and Extend Performance overall in the same benchmark. 
• • Reducto’s r-1 mini outperformed Extend Performance at 95% lower cost and LlamaParse Agentic at 92% lower cost.

There are more accurate options for particularly demanding documents, such as r-1. But r-1 mini was the lowest-priced full-layout parser we evaluated, and it beat every tested entry-level layout and parsing configuration overall.

Here’s what we observed across the tested configurations in terms of breadth of features:

Our dedicated benchmark post will cover the methodology, individual results, and differences across document categories in more detail.

r-1 remains the choice for frontier performance

We’re also making r-1 generally available. r-1 is our full-featured flagship model, delivering our highest accuracy and lower latency than r-1 mini, with advanced capabilities such as word-level citations included.

r-1’s accuracy is especially valuable on demanding inputs such as charts, diagrams, chemical equations, degraded scans, and mixed-language documents. You can expect to see 30-50% better latency with r-1 compared to r-1 mini, making it a great choice for synchronous or live-use cases.

During the r-1 preview, customers saw substantial improvements on edge cases their previous pipelines struggled with, while reducing processing costs.

If you’re deciding between r-1 and r-1 mini, we’re happy to help you benchmark both models on your own documents. Teams switching from another document processing provider can also receive up to $5,000 in migration credits.

Everyone can try Reducto today

We’re opening signup to personal email addresses, so independent developers, researchers, and anyone exploring a new idea can get started with free credits included. Happy parsing.

Try r-1 mini →

Written by: Adel Wu, Yifei Hu, Janna Lu, and Archana Ahlawat