Corrected article
DocsRouter has not published an OCR benchmark dataset, ELO scores, or a verified accuracy ranking. Earlier ranking claims on this page were unsupported and have been removed.
There is no universal best OCR model. Results depend on language, handwriting, layout, scan quality, field definitions, and the errors your workflow can tolerate.
A useful evaluation process
- Build a representative sample and reviewed ground truth.
- Define field-level success criteria before testing.
- Run the same inputs and instructions against supported model IDs from
GET /v1/models. - Record extraction errors, provider cost, and observed processing time.
- Repeat after meaningful provider or prompt changes.
DocsRouter's model descriptions and static routing policies are practical starting points, not measured leaderboards. quality and accuracy currently select google/gemini-2.5-pro; speed and balanced select google/gemini-2.5-flash; cost selects openai/gpt-4o-mini.
Those labels do not promise a measured accuracy or latency level, and DocsRouter does not automatically fall back to another model.