DEMO ARCBeat 6 of 6 · Compare (the wedge)
All screens
Same invoice through Mistral OCR, Qwen-VL, Pixtral, and our LayoutLMv3 specialist. We pick the right tool per supplier. For FGF’s volume, the cost difference is six figures a year.
5. Memory (the moat at runtime)final beat · close with honest tradeoff + roadmap

Compare extractors

Same invoice, four extractors, side-by-side. Disagreements get flagged. Pick the winner manually or trust the consensus.

FIELDS

6 extracted

consensus + per-field conf

API

Mistral OCR 3

850 ms · $0.00200/page

self-hosted (A100)

Qwen2.5-VL-72B

1450 ms · $0.00012/page

self-hosted (A100)

Pixtral-12B

920 ms · $0.00008/page

self-hosted (CPU)

LayoutLMv3 (specialized)

240 ms · $0.00001/page

FieldMistral OCR 3Qwen2.5-VL-72BPixtral-12BLayoutLMv3 (specialized)Consensus
supplier_name
Maple Leaf Flour Coconf 0.99
Maple Leaf Flour Coconf 0.98
Maple Leaf Flour Coconf 0.97
Maple Leaf Flour Coconf 1.00
Maple Leaf Flour Co
invoice_number
INV-2026-04812conf 0.99
INV-2026-04812conf 0.99
INV-2026-04812conf 0.97
INV-2026-04812conf 1.00
INV-2026-04812
invoice_date
2026-05-14conf 0.97
2026-05-14conf 0.98
2026-05-14conf 0.99
2026-05-14conf 1.00
2026-05-14
subtotal
3245.80conf 0.95
3245.80conf 0.95
3245.80conf 0.94
3245.80conf 1.00
3245.80
tax
162.29conf 0.97
152.29conf 0.62⚠ disagrees
162.29conf 0.93
162.29conf 1.00
162.29
total
3408.09conf 0.99
3398.09conf 0.71⚠ disagrees
3408.09conf 0.95
3408.09conf 1.00
3408.09

Cost / latency / agreement

For this invoice across all four extractors. Run on the corpus to see population-level numbers.

ExtractorHostingLatencyCost / pageAgreement w/ consensusNotes
Mistral OCR 3API850 ms$0.002006/6 (100%)industry baseline
Qwen2.5-VL-72Bself-hosted (A100)1450 ms$0.000124/6 (67%)open weights
Pixtral-12Bself-hosted (A100)920 ms$0.000086/6 (100%)Mistral open
LayoutLMv3 (specialized)self-hosted (CPU)240 ms$0.000016/6 (100%)fine-tuned for Maple Leaf

The cost wedge

For FGF's 200k invoices/year, the difference between extractor choices compounds fast.

StrategyCost / pageAnnual (200k pages)Trade-off
Nanonets / Rossum (flat API)$0.10 – $0.40$20,000 – $80,000one model, no specialization
Mistral OCR 3$0.002$400API dependency, no fine-tune
Qwen / Pixtral (self-hosted)$0.00008 – $0.00012$16 – $24infra setup, open weights
Per-supplier specialist (LayoutLM)≈ $0$0requires per-supplier training; only worth it at volume

The system in the registry routes each supplier to the cheapest extractor that still hits the confidence floor — most volume converges on the bottom row over time.