Evals

Every model + prompt combination is run against a golden dataset of 30 synthetic invoices with exactly known ground truth. Scores are field-level exact-match accuracy after normalization (dates → ISO, amounts → numbers, currency → ISO 4217).

Accuracy by field

Field
haiku-4-5
prompt v2 · 30 docs
vendor_name100%
invoice_number100%
invoice_date100%
due_date100%
currency100%
subtotal97%
tax_amount100%
total100%
line_items97%
docs 100% correct97%
avg cost / doc$0.0047
avg latency3192 ms

All runs

WhenModelPromptRuntimeDocsDocs 100% correctAvg cost/docAvg latency
2026-07-10claude-haiku-4-5v2api3097%$0.00473192 ms