alphaxiv used DeepSeek-OCR to extract datasets from 500k+ arXiv papers for $1000 (vs $7500 with Mistral OCR)

TL;DR

alphaxiv (the arXiv reading platform) ran DeepSeek-OCR over the tables and charts of 500k+ AI arXiv papers to harvest every benchmark / dataset mention into a searchable index. Their headline number: the full corpus pass cost ~ $1,000, versus a quoted ~$ 7,500 if they had used Mistral OCR — a ~7.5× cost ratio for the same task. The tweet positions DeepSeek-OCR’s “contexts optical compression” trick (vision tokens compressing text 7–20×) as not just a research curiosity but a production-grade cost lever for large-scale scientific-document parsing.

Key claims

DeepSeek-OCR was used to extract every dataset/benchmark mention from tables and charts across 500k+ AI arXiv papers [tweet body].
The total cost was ~ $1,000 with DeepSeek-OCR vs an estimated ~$ 7,500 with Mistral OCR — a ~7.5× cost reduction at constant task [tweet body].
The output is a searchable index of trending benchmarks and dataset discovery surface, hosted on the alphaxiv platform [tweet body].

Method

The tweet is a one-shot announcement with no methods detail beyond the model choice (DeepSeek-OCR) and the corpus size (500k+ arXiv papers). The implied pipeline: render each paper’s table / chart regions, run DeepSeek-OCR for structured extraction, parse outputs into a benchmark-dataset relation table, then index for search. The cost comparison is presumably token-rate-vs-token-rate at advertised API prices; no per-paper latency or accuracy numbers are given in the tweet.

Results

The only quantitative results in the tweet are the two dollar figures: ~ $1,000 with DeepSeek-OCR vs ~$ 7,500 with Mistral OCR for the same 500k-paper extraction task. No accuracy, recall, or comparison-against-ground-truth numbers are provided, and the tweet does not say which model produced higher-quality extractions — only which was cheaper.

Why it’s interesting

This is the first concrete production-scale cost number filed in the wiki for DeepSeek-OCR — until now DeepSeek-OCR: Contexts Optical Compression and Karpathy: pixels may be better LLM inputs than text tokens (DeepSeek-OCR springboard) discussed the model and its “pixels as compressed text tokens” thesis in research terms, but neither attached a real downstream-pipeline dollar figure. It pairs naturally with ArxivQA: Training Retrieval Agents for arXiv Search — the same group’s separate effort to make arXiv content retrievable — suggesting alphaxiv is building a multi-stage pipeline where cheap vision-OCR underpins higher-level retrieval agents. The unverified claim is the headline 7.5× cost ratio: at constant accuracy this would meaningfully shift cost-curves for any large-scale scientific-document parsing workload, but the tweet provides no accuracy comparison to anchor that.