What is the best model for data extraction?
As of September 7, 2026, the cheapest live-measured option within 90% of top quality on information extraction is granite-4.0-h-micro at $0.0144 per 1,000 requests — 0.904 measured quality vs 0.977 at the top, with a 252× price spread across the measured frontier (frontier v5).
re-measured weekly · how these numbers are made
Extraction — pulling fields out of documents, emails, and records into structured JSON — is scored field-by-field against known references, so the numbers below are deterministic, not judged. It is also the workload where reference-free LLM judging fails measurably (judges cannot see omissions), which is why extraction claims anywhere should be treated with suspicion unless the scoring method is stated.
The measured frontier
Every row is a live measurement on Potion's held-out information extraction suite — same items, same scoring, per option. Names that are part of the product are withheld; their numbers are not.
| option | vendor | measured quality | $ / 1K requests | p95 latency |
|---|---|---|---|---|
| granite-4.0-h-micro | ibm-granite | 0.904 | $0.0144 | 3105 ms |
| Potion's routed pick (name withheld) | — | 0.945 | $0.0208 | 12858 ms |
| mistral-nemo | mistralai | 0.917 | $0.0244 | 3848 ms |
| deepseek-chat-v3.1 | deepseek | 0.950 | $0.1824 | 11996 ms |
| inkling-small | thinkingmachines | 0.977 | $0.5466 | 4649 ms |
| claude-opus-4.6 | anthropic | 0.962 | $3.62 | 2723 ms |
frontier v5 · measured 2026-09-07 · live provider calls only — simulated evidence never appears on this page
Head to head
granite-4.0-h-micro vs mistral-nemodeepseek-chat-v3.1 vs granite-4.0-h-microgranite-4.0-h-micro vs inkling-smallclaude-opus-4.6 vs granite-4.0-h-microdeepseek-chat-v3.1 vs mistral-nemoinkling-small vs mistral-nemoclaude-opus-4.6 vs mistral-nemodeepseek-chat-v3.1 vs inkling-smallclaude-opus-4.6 vs deepseek-chat-v3.1claude-opus-4.6 vs inkling-small
Questions
How is extraction quality measured?
Field-match against a known reference: every extracted field either matches or it does not, with fractional credit per document. No judge model is involved; a missing field cannot hide.
Can a small model really do production extraction?
On measured suites, the gap between the best cheap model and the best frontier model is often inside the confidence interval. Real documents are messier than any suite — which is why the measured answer is a starting point and customer-traffic verification is the finish line.
What about PDFs, scans, and images?
Vision extraction is measured separately, on rendered documents. The frontier below is text extraction; vision numbers appear in the weekly issues as they are re-measured.
Potion routes each request to the cheapest option measured at your quality bar — with a receipt on every answer and this page's evidence behind every pick. Get an API key or read the docs.