We Audited 30 Days of Our Own Billing. Here Is What It Shows.

"0% markup" is easy to print on a pricing page. It is harder to check. So we checked it the only way that means anything: take every paid text call we served in 30 days, recompute its cost from the provider list price and the token counts, and compare that to what the billing system actually recorded.

The audit

The largest single-model deviation was 1.72% (GLM-5.1), which comes from rounding per-call costs to six decimals. Everything else is under half a percent. If there were a markup, it would show up here — it does not.

Where the money actually goes

This is the part most cost discussions get wrong. Output tokens are priced 2-5x higher than input tokens, so people assume output dominates the bill. In this traffic it does not: input is 95.8% of the cost, because the workload sends large prompts and receives short answers — an average of about 40K input tokens per call against 815 output tokens.

Three practical consequences

1. Optimise the prompt before the output

If input is ~96% of the bill, trimming retrieved context, adding prompt caching or capping document size saves far more than shortening answers. Shortening output by 20% saves under 1% of the bill in this mix.

2. Model choice is an 80x lever

Realized blended cost per 1M tokens ranged from $0.0806 (deepseek-v4-flash, 5,377 calls) to $6.46 (kimi-k2.7-code-highspeed, 37 calls) — an 80x spread for the same token volume. Flagship models earn their price on hard tasks; they do not need to run on easy ones.

3. Price per token is a starting point, not an estimate

A call's cost is (input tokens × input price) + (output tokens × output price). At ~40K input tokens per call, even a cheap model produces meaningful per-call costs — which is why the call-level distribution is a better planning tool than the rate card.

Reproduce it

Every number here can be recomputed from the public price feed plus your own token counts:

cost = input_tokens * input_price_per_1m / 1e6
     + output_tokens * output_price_per_1m / 1e6

curl -s https://aiapi-pro.com/china-ai-leaderboard.json | jq '.categories[] | .models[] | {name, price_usd}'

FAQ

Does NovAI add a markup on model prices?

No. Auditing 30 days of billing across 58 models with published list prices, the realized cost was $188.74 against a list-implied cost of $188.74 — a 0.005% deviation, which is rounding. The gateway bills at provider list price with a 1.0 multiplier.

Are input or output tokens the bigger cost?

In this traffic, input. 556.2M input tokens versus 6.1M output tokens meant input was 95.8% of the bill and output 4.2% — even though output tokens are priced 2-5x higher per token.

Why is input so dominant?

Agent and retrieval workloads resend a large prompt (averaging about 40K input tokens per call) and receive a short answer. The pattern is common for coding agents, RAG and chat-with-documents products.

How much does model choice change the bill?

By up to 80x. In the same 30 days, the realized blended cost ranged from about $0.08 per 1M tokens on deepseek-v4-flash to $6.46 per 1M on kimi-k2.7-code-highspeed.

Audit window: 30 days ending 2026-10-22. 58 text models with published list prices; media and free-tier calls excluded. Realized cost from billing records; list-implied cost from token counts (input/output) multiplied by list prices. Prices: /china-ai-leaderboard.json.