ParseBench-2 / OCRia-1.5

ParseBench-2 and OCRia-1.5 are the document-parsing benchmark and model family LlamaIndex released together on 12 November 2026. The official blog post Introducing ParseBench-2 (dated Nov 12, 2026, bylined Gauthami Polasani and Pierre-Loic Doulcet) says: “Today, we’re proud to raise the standard for document parsing benchmarks with ParseBench-2, our most comprehensive document parsing benchmark yet.” Founder @jerryjliu0 posted at 2026-11-12T16:46:42Z (00:46 Beijing time on 13 November): “Today we’re introducing ParseBench-2 - our most ambitious document parsing benchmark to date.”

Officially, ParseBench-2’s five main changes are: deterministic metrics calibrated to human-defined severity; a corpus expanded from about 2,100 cases in ParseBench-1 to about 4,300; corrected and clarified annotations on the original five dimensions; complete value-by-value checks on 1,025 charts (the post cites 36,958 values); and forms as a sixth scoring dimension. The company says it defined 330+ error types, classified as critical, major, minor, or harmless, and used them to calibrate metrics across text, tables, formatting, charts, layout, and forms. The post contrasts a 459-cell table: under ParseBench-1, a swapped number cost about 0.13 points while a correct table returned as Markdown could score zero; under ParseBench-2, the wrong number costs 10 points and the Markdown output deducts 0. Those figures and examples are the company’s; this page does not reproduce them independently. Officially, the datasets, evaluation code, and paper are public; the documented clone path is git clone https://github.com/run-llama/ParseBench.

The same article says: “In conjunction with ParseBench-2, we’re introducing OCRia-1.5, our new family of document parsing models powering LlamaParse’s Cost Effective, Agentic and Agentic Plus modes.” Officially, the models improved across dimensions, especially tables, forms, and visual layout, with a fuller write-up promised in a follow-up post. The blog also publishes a company-run ParseBench-2 ranking: LlamaParse Agentic Plus 89.8 overall, Agentic 86.0, and Cost Effective 84.6, plus comparisons with other parsers and directly called vision models, and per-page cost figures. Those scores are reported by LlamaIndex on its own benchmark and are not independently verified here. Do not describe ParseBench-2 as OmniParseBench, which appears in other entries, or as a routine version bump of ParseBench-1 that was never formally launched; and do not describe OCRia-1.5 as the earlier OCRia-1 line without this 12 November announcement.

This entry is a catch-up for the 12 November official launch. Jerry Liu’s post came just after the 00:32 scan, and the 04:32, 08:32, and 12:32 passes did not store the two names.

Key points:

• Official blog: https://www.llamaindex.ai/blog/introducing-parsebench-v2 ; introduction post: https://x.com/jerryjliu0/status/2120921239619612921 ; benchmark repository: https://github.com/run-llama/ParseBench ;
• These are two concrete LlamaIndex names (a new benchmark and a new model family). Do not confuse them with ParseBench-1, OmniParseBench, or the existing entry for Chandra 2.7. A fuller official description of OCRia-1.5 is still promised in a later post.

Why it matters

These are formal names LlamaIndex published on 12 November 2026 in an official blog post and a founder post: ParseBench-2 is the upgraded document-parsing benchmark, and OCRia-1.5 is the new model family that officially powers three LlamaParse pricing tiers. The introduction was at about 16:46 UTC (00:46 Beijing time on 13 November); later scans the same day did not store the names, and this pass adds them.

This page only restates the official blog and introduction post. Leaderboard scores, per-page costs, and the 330+ error types are company statements and are not independently verified here.

Limits

Official materials say a detailed write-up of OCRia-1.5 will come later; do not treat this page as a complete technical description of that model family. ParseBench-2 is published by LlamaIndex, and LlamaParse configurations lead the company’s own ranking; cite scores with that source and interest in mind. Do not treat this entry as an independent verdict on any parser or on downstream extraction quality.

1. Read the official blog at https://www.llamaindex.ai/blog/introducing-parsebench-v2 , and cross-check the founder post at https://x.com/jerryjliu0/status/2120921239619612921 and the repository at https://github.com/run-llama/ParseBench ;
2. Follow the official instructions to clone the repository and run the evaluation command; when using LlamaParse, note the official statement that OCRia-1.5 powers the Cost Effective, Agentic, and Agentic Plus modes, and watch for the promised follow-up model post;
3. When citing the 4.3k corpus, 330+ error types, six dimensions, or leaderboard scores, attribute them to the 12 November 2026 official blog and recheck the upstream source;
4. Do not describe ParseBench-2 / OCRia-1.5 as ParseBench-1, OmniParseBench, or the existing entry for Chandra 2.7; this page only restates official materials — recheck the upstream sources before citing them externally.