Pricing pages quote per million tokens. That number does not tell you what a call costs, because calls vary by three orders of magnitude. So we measured our own traffic: every paid call served in the last 30 days, aggregated. No user data — just the cost of each request.
| Percentile | Cost per call | What it looks like |
|---|---|---|
| p10 | $0.00003 | A short prompt on the cheapest model |
| p25 | $0.00020 | Chat turn, small context |
| p50 (median) | $0.0048 | Typical assistant turn with retrieved context |
| p75 | $0.0133 | Longer prompt, more output |
| p90 | $0.0279 | Large context or a pricier model |
| p99 | $0.3846 | Long-context flagship call |
Mean: $0.0175. The mean sits 3.6x above the median because the tail is heavy — a small number of long-context or flagship calls carry a large share of the total. If you plan with the mean, you plan for the tail; if you plan with the median, you plan for the typical case.
The free tier is 10 free-model calls per day. At the median call cost that is about $0.05 of usage per day — roughly $1.45 a month at typical paid rates.
That is not a criticism of free tiers; it is what they are for. Ten calls a day is enough to:
It is not enough to run a production workload — and pretending otherwise would be the wrong pitch. The honest use of a free tier is measurement.
| Workload | At median cost | At mean cost |
|---|---|---|
| 20 calls/day, 30 days | $2.90 | $10.50 |
| 50 calls/day, 30 days | $7.20 | $26.20 |
| 100 calls/day, 30 days | $14.50 | $52.40 |
| 500 calls/day, 30 days | $72.50 | $262.00 |
Top-ups start at $1 on card ($5 on USDT / Alipay), and the first one carries a 25% bonus. In call terms:
| Top-up | Calls at median | Calls at mean |
|---|---|---|
| $5 | ~1,040 | ~286 |
| $5 with first-top-up bonus ($6.25) | ~1,300 | ~357 |
| $25 with bonus ($31.25) | ~6,510 | ~1,790 |
If your product has one feature that sends 200K-token documents to a flagship model, that feature will not cost "the average" — it will cost like the p99. Two practical rules:
# cost of one call = (input_tokens * input_price + output_tokens * output_price) / 1e6
curl https://aiapi-pro.com/china-ai-leaderboard.json | jq '.categories[] | select(.category=="Chat / LLM") | .models[] | {name, price_usd}'Across 13,807 paid calls in a 30-day window, the median call cost $0.0048 and the mean cost $0.0175. The mean is higher because a small number of long-context or flagship-model calls dominate the tail: p90 is $0.028 and p99 is $0.385.
At the median call cost, 10 free calls per day is about $0.05 per day — enough to measure latency, output quality and integration behaviour, not enough to run a production workload.
About 1,000 calls at the median cost, or about 285 calls at the mean cost. On a first top-up the 25% bonus raises that to roughly 1,250 and 355 calls respectively.
When your workload needs more than 10 calls per day or 5 requests per minute, or when you move from evaluation to production. At the mean call cost, 100 calls per day is about $52 per 30-day month.
10 free calls a day, no card. When the math says top up, the first one carries a 25% bonus.
Get a free API keyDistribution measured from NovAI usage logs over a 30-day window ending 2026-10-22 (13,807 paid calls; free-tier calls excluded). Per-model list prices at /china-ai-leaderboard.json.