RAG vs Long-Context Cost Calculator

Both approaches answer the same question with the same generation model; the only difference is how many tokens you feed it. Long context sends the entire corpus every query; RAG sends only the top-k retrieved chunks. The calculator prices each side from a dated static table of vendor per-token rates — no network calls — so results are deterministic.

Write C for corpus tokens, q for question tokens, a for answer tokens, k for top-k, s for chunk size, and per-token prices pin, pout.

• Long context, uncached: per query = (C + q)·pin + a·pout. Monthly = queries × that.
• Long context, cached: per query = C·pcr + q·pin + a·pout where pcr = 0.1·pin (Anthropic cache-read). The cache is re-written once per 5-minute window, so writes per month = min(queries, 8640) each costing C·pcw (pcw = 1.25·pin). If caching ends up more expensive than uncached — which happens at low volume — the tool uses the uncached price, because a rational operator would turn caching off.
• Long-context premium: Google and Anthropic reprice the whole request once input exceeds 200,000 tokens, so a large corpus is billed entirely at the higher rate.
• RAG: a one-time embedding of C·pemb; storage for ceil(C/s) vectors at ${0.33}/GB-month (binary GiB, 4 bytes per dimension); and per query q·pemb + (min(k, chunks)·s + q)·pin + a·pout. Retrieval is capped by the corpus — RAG never sends more than exists.

The break-evenis a corpus size, not a query count. Because RAG's per-query cost flattens once the corpus is bigger than one retrieval (it always fetches the same top-k), while long context grows with corpus size, the two cross at about k·stokens. Below that corpus, long context is cheaper and simpler; above it, RAG pulls ahead and the gap widens with corpus size. Every rate cites the vendor page in the source module and is cross-checked by an independent per-1,000-token arithmetic path (the same technique the site's tax calculator uses against IRD Table 01).

• Anthropic — Claude model pricing, prompt caching, and 1M-context beta rates
• OpenAI — API and embedding (text-embedding-3-small / -large) pricing
• Google — Gemini API pricing (incl. the >200K long-context tier)
• Pinecone — serverless vector storage pricing ($/GB-month)
• Central Bank of Sri Lanka — indicative USD→LKR exchange rate

Prices were last cross-checked against the vendor pages on 2026-07-18. This is a cost comparison only — it does not model answer quality or latency, which depend on your data and setup.