What is the best model for RAG answers?
As of August 24, 2026, the cheapest live-measured option within 90% of top quality on rag question answering is Potion's routed pick (name withheld) at $0.0058 per 1,000 requests — 0.907 measured quality vs 0.982 at the top, with a 16× price spread across the measured frontier (frontier v4).
re-measured weekly · how these numbers are made
In retrieval-augmented answering, the model’s job narrows to faithful synthesis over supplied context — a different skill from open-ended knowledge, and one where paying for a frontier model’s world knowledge is often paying twice. The measured suite scores answers against references derived from the supplied documents.
The measured frontier
Every row is a live measurement on Potion's held-out rag question answering 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 |
|---|---|---|---|---|
| Potion's routed pick (name withheld) | — | 0.907 | $0.0058 | 4864 ms |
| ling-3.0-flash | inclusionai | 0.963 | $0.0067 | 1660 ms |
| inkling-small | thinkingmachines | 0.982 | $0.0937 | 1653 ms |
frontier v4 · measured 2026-08-24 · live provider calls only — simulated evidence never appears on this page
Head to head
inkling-small vs ling-3.0-flash
Questions
Does a bigger model help when the context contains the answer?
Less than expected: the measured gap narrows sharply when answers must come from supplied context. Faithfulness, not knowledge, dominates — and is measurable.
How is faithfulness measured?
Against document-derived references: an answer scores on agreement with what the supplied context supports, so a fluent hallucination scores as the miss it is.
Should retrieval and answering use the same model?
They are different workloads; route them separately. Embedding and answering frontiers are measured independently.
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.