Real-call cost benchmarks: support, RAG and batch
Short answer: A support bot answering 2,000-token messages costs fractions of a cent per call on small models; a RAG answer with a 20,000-token context costs more; batch summaries are roughly half price where a provider publishes a batch multiplier. The tables below price all three workloads at official list rates.
Three workloads that teams actually budget for, priced through the same models with the arithmetic the calculator runs and the rates in src/prices.json. Every number on this page is generated by build.py from that file at build time — none of it is typed in by hand, and none of it is a measurement of a real customer.
How to read the tables
Per call, per 1,000 calls and per month
“Per call” is one message, one question or one document. “Per 1,000 calls” scales it for a sanity check. “Per month” applies the volume in the label above each table. “Min price” is the calculator's minimum subscription price at the margin shown: cost per user (or account) a month ÷ (1 − margin). All figures are USD at list price.
cost per call = (uncached input × input rate + cached input × cached rate + output × output rate) ÷ 1,000,000
Assumptions are labelled, not hidden
The token counts under each heading are modelling assumptions, flagged as such, not measurements of a real customer. The rates they are multiplied by are the hand-checked list prices in the site's data files.
The same formula is implemented twice — cost() in build.py and price() in src/app.js — and test/prices.test.mjs checks that both give the same answer for every verified model. Batch rows multiply the standard cost by the batch multiplier in src/specs.json (0.5 for OpenAI, Anthropic and Google, 0.8 for the xAI Grok models that publish one).
Support bot
Assumption, not a measurement: 1,000 users × 20 messages a month = 20,000 calls; 2,000 input and 500 output tokens a call; standard list price; 80% target gross margin. The token shape is the calculator's “Support bot” preset: a system prompt, two turns of history and a short reply. Nothing here assumes your users write the same number of words.
| Model | Per call | Per 1,000 calls | Per month | Min price / user at 80% |
|---|---|---|---|---|
| GPT-5.4 mini | $0.00375 | $3.75 | $75.00 | $0.3750 |
| Claude Sonnet 5.5 | $0.0090 | $9.00 | $180 | $0.9000 |
| Gemini 3.8 Flash | $0.00337 | $3.38 | $67.50 | $0.3375 |
| Grok 4.3 | $0.00375 | $3.75 | $75.00 | $0.3750 |
| DeepSeek Flash (V4.1) | $0.0012 | $1.20 | $24.00 | $0.1200 |
| Gemini 2.5 Flash-Lite | $0.0004 | $0.4000 | $8.00 | $0.0400 |
The cheapest model for this workload in the current snapshot is Ministral 3 3B at $0.00025 a call, $5.00 a month for the volume above.
RAG document Q&A
Assumption, not a measurement: 500 users × 20 questions a month = 10,000 calls; 20,000 input and 1,000 output tokens a call; standard list price; 80% target gross margin. The token shape is the calculator's “RAG answer” preset: a retrieved context of about 20,000 tokens and a 1,000-token answer. Retrieved passages, the question and the instructions are all billed as input.
| Model | Per call | Per 1,000 calls | Per month | Min price / user at 80% |
|---|---|---|---|---|
| GPT-5.4 mini | $0.0195 | $19.50 | $195 | $1.95 |
| Claude Sonnet 5.5 | $0.0500 | $50.00 | $500 | $5.00 |
| Gemini 3.8 Flash | $0.0187 | $18.75 | $188 | $1.88 |
| Grok 4.3 | $0.0275 | $27.50 | $275 | $2.75 |
| DeepSeek Flash (V4.1) | $0.0072 | $7.20 | $72.00 | $0.7200 |
| Gemini 2.5 Flash-Lite | $0.0024 | $2.40 | $24.00 | $0.2400 |
The cheapest model for this workload in the current snapshot is Ministral 3 3B at $0.0021 a call, $21.00 a month for the volume above.
Batch summarisation
Assumption, not a measurement: 1,000 accounts × 200 documents a month = 200,000 documents; 3,000 input and 500 output tokens a call; standard list price; the provider batch multiplier applied; 80% target gross margin. A single 3,000-token document in, a 500-token summary out, run through the Batch API. The batch multiplier from src/specs.json is applied where the provider publishes one; where it does not, the row says so instead of assuming a discount.
| Model | Per call (batch) | Per 1,000 calls | Per month | Min price / account at 80% |
|---|---|---|---|---|
| GPT-5.4 mini | $0.00225 | $2.25 | $450 | $2.25 |
| Claude Sonnet 5.5 | $0.0055 | $5.50 | $1,100 | $5.50 |
| Gemini 3.8 Flash | $0.00206 | $2.06 | $412 | $2.06 |
| Grok 4.3 | $0.0040 | $4.00 | $800 | $4.00 |
| DeepSeek Flash (V4.1) | no batch rate published for this model | |||
| Gemini 2.5 Flash-Lite | $0.00025 | $0.2500 | $50.00 | $0.2500 |
The cheapest model for this workload in the current snapshot is Ministral 3 3B at $0.000175 a call, $35.00 a month for the volume above.
What moves these numbers
- Output tokens. Output is priced several times higher than input on most models, so answer length moves a bill more than prompt length. A support reply that grows from 500 to 1,000 tokens costs more than doubling the prompt.
- Retrieved context. Every retrieved passage is billed as input on every request. Trimming five 4,000-token chunks to the three that matter is a bigger saving than switching model tiers.
- Caching. A fixed system prompt or a shared document prefix can be billed at the cached-input rate; the caching guide shows the break-even on a 20K prefix.
- Batch. Asynchronous work is roughly half price where a batch rate exists, which is why the third table applies it. It does not make interactive work cheaper.
Check it with your own token counts
These are modelling assumptions, not benchmarks of quality or of your traffic. Log 50–100 real requests, average the input and output token counts the API returns, and put those numbers into the calculator with your own volume. The token guide explains what counts as input, and the methodology page lists where every rate comes from. For measured scores rather than cost, see the LLM benchmark comparison; for a use-case recommendation, use the model finder.
Questions
Are these costs measured from real customers or modelled assumptions?
Modelled. Each table states its token counts and monthly volume, and the rates come from the official prices in <code>src/prices.json</code>. Nothing here is a sample of a real customer's invoice; it is the arithmetic a budget should start from before you replace the assumptions with your own logs.
Why does each workload use a different monthly volume?
Because each scenario is defined for a different unit: 20 messages per user a month for support, 20 questions per user for RAG, and 200 documents per account for batch summarisation. The per-month column scales with that volume, so compare scenarios on cost per call and cost per user, not only on the monthly total.
Why is the batch multiplier applied to one table only?
The first two scenarios are interactive — a user is waiting — so they pay the synchronous rate. Summarisation is asynchronous, so the table applies the batch multiplier from <code>src/specs.json</code> where a provider publishes one and says so where it does not. Mixing the discount into interactive work would understate the bill.
How do I turn these numbers into a subscription price?
Take the per-user monthly cost and divide by (1 − target gross margin). The <a href="/ai-saas-unit-economics-calculator">unit-economics calculator</a> does it for your own users, calls and token counts, and the <a href="/guides/how-to-price-an-ai-saas">pricing guide</a> covers the tail users.
Next step → Turn a monthly bill into a per-user price
Price data last checked 2026-11-18 against official sources · how we verify
Price-change alerts
One short email when a provider changes a price or a promo ends. No spam, unsubscribe any time.