LlamaIndex co-founder Jerry Liu says Mistral OCR 4 scored 60.7 on ParseBench, trailing MinerU but beating AWS Textract

We've provided some updated results on Mistral OCR that make use of the annotation feature for charts.

The overall score is ahead of GPT-5.5 and just behind Gemini 3.1 Pro, which is quite impressive for a model at its price range.

It does a great job on content faithfulness, semantic formatting, and visual grounding. It does an ok job on tables and an ok job (though to be fair, non-zero) on charts.

There's been some great advancements on visual understanding capabilities lately. See screenshot below for performance, will update the main benchmark page soon: https://www.parsebench.ai/

@jerryjliu0 Thanks a lot for evaluating the model! By the way, we provide an annotation feature for charts and other types of images - it should considerably boost the Charts score as the default model is optimized for OCR only 👀

Pricing is a bit different, 5$/1k instead of 4$.

@pandoraxtw will take a look

@jerryjliu0 Can you add https://x.com/Baidu_Inc/status/2069358973753729165 to the leaderboard please? (It's based on deepseek ocr)

@jerryjliu0 finally

@jerryjliu0 Good benchmark. Real-world OCR is mostly blurry faxes, handwritten margins, merged table cells. How does it do on those?

@jerryjliu0 I like mistral ocr, havent tried v4 yet. I love their relatively high concurrent-requests limit. Especially if you're on the PAYG plan. Personally I like using it purely for the OCR step and then I use 3.1 Flash-Lite for the extraction 👀

Semantic formatting is where most OCR pipelines quietly fail in production. In lending docs - promissory notes, HELOCs, appraisals - a misread superscript on a rate or a missed strikethrough on a clause is a real liability. Curious how ParseBench handles tables with merged cells and footnotes, that's the hardest case most financial systems hit consistently.

@jerryjliu0 Do you have a Pareto plot?

@fahdmirza Thanks for the video!

@jerryjliu0 interesting to see it trade blows with models that cost way more per page

the semantic formatting wins are the sleeper feature

@Arcturus_f yes we've evaluated! will post the notes soon

@jerryjliu0 Another thing that could help for tables, is to toggle html mode, we support different table formats.

@jerryjliu0 the jump in performance at that price point is wild

@jerryjliu0 @jerryjliu0 Price-point competitiveness on semantic formatting is the real story — most enterprise document workflows fail not on raw accuracy but on whether the output is structured enough to act on. Mistral is hitting the threshold that matters.