IDP Leaderboard: Sonnet 4.6 Matches Opus 4.6 on Real-World Document Understanding

One-sentence takeaway

On the open document AI leaderboard, Claude Sonnet 4.6 scores 80.7 overall, slightly above Opus 4.6's 80.4, making Sonnet a good choice for offloading OCR, table extraction, layout understanding, and key information extraction from Opus; still watch for content moderation false positives on archived scans.

Test environment

• Model: Leaderboard rank #8 Claude Sonnet 4.6, rank #9 Claude Opus 4.6, rank #16 Claude Haiku 4.5.

• Task: Intelligent Document Processing, covering OCR, table extraction, key information extraction, and visual Q&A; the overall score is the mean of benchmark subscores.

• Sub-benchmarks: OlmOCR, OmniDoc, IDP.

• Scale: The Reddit sync post cites 16 models and 9000+ real documents; at collection time the main table had expanded to 26 models. Treat the leaderboard's current numbers as authoritative.

• Artifacts: The site provides Code, Datasets, and Results Explorer; methodology links to GitHub.

Inputs/configuration

The homepage does not list the full prompt, temperature, effort, or whether thinking is enabled for each Claude model. For verification, open Results Explorer for per-document outputs and check the GitHub methodology for the harness.

Results data

Main table at collection time (Overall / OlmOCR / OmniDoc / IDP):

The Reddit sync post at the time listed Sonnet 80.8 / Opus 80.3 / Haiku 69.6, and said extraction-task radar charts were nearly identical; Sonnet cost about $24/1K pages and Opus about $40/1K pages. Main table numbers have since been tweaked slightly; cite https://www.idp-leaderboard.org/ when quoting.

The sync post also noted that old newspaper scans, textbook pages, and historical documents sometimes trigger stricter Claude content moderation, mainly on OlmOCR and OmniDoc.

Conclusion

For document extraction, tables, and layout understanding, Sonnet 4.6 can replace Opus 4.6 as the default model; Haiku 4.5 is clearly behind. If the use case is specialized OCR, the current top entry is Nanonets OCR-3, not general-purpose Claude. Archived and historical scans require separate testing of moderation block rates.

Limitations

• This is a document understanding leaderboard; do not extrapolate to SWE, computer use, or open-ended reasoning.

• Reddit's older numbers differ from the main table at collection time by 0.1–1.6 points; cite the page snapshot at the time of reference.

• The homepage does not publish Claude's full sampling configuration; cost figures come from the Reddit sync post, not leaderboard main table fields.

• Content moderation failures count toward relevant subscores; they do not mean the model "couldn't understand" the page.

Reproduction steps

1. Download the public datasets and evaluation code from the site, and fix the model ID claude-sonnet-4-6 and comparison models.

2. Run OlmOCR, OmniDoc, and IDP per the GitHub methodology, and save per-page predictions.

3. Report extraction accuracy, moderation block rate, and cost per thousand pages separately.

4. Spot-check failed pages in Results Explorer to distinguish recognition errors from safety filtering.