How to verify what an LLM gateway actually charges you

Inference pricing has a trust problem. List prices differ from billed prices, "savings" percentages are computed against convenient baselines, and cached tokens, the thing agents mostly pay for, are usually invisible in the marketing. The fix is not better marketing. It is per-request verifiability.

The baseline trick to watch for

A provider advertising "40% savings from caching" is usually comparing their input bill against an imaginary uncached one, not telling you what a session costs. Output tokens, which never cache, often dominate real spend once caching is working, so total savings land well below the headline number. The only comparison that means anything is dollars for your workload: your token mix, your cache-hit rate, your hours of the day.

What above.dev puts in every response

Every API response carries enough headers to reconstruct its exact charge:

• x-cost-usd and x-cost-micro: the precise amount deducted for this request.
• x-input-cache-hit-tokens and x-input-cache-miss-tokens: how your input split between the cheap and full rate.
• x-input-cache-hit-rate: the fraction of input served from cache, for example 0.9553.

Multiply the token counts by the published per-million rates for the model and you get the same number the header shows. If you ever cannot, that is a bug and we want to know.

A five-minute audit for any provider

1. Run one real agent session, not a hello-world request. Trivial prompts cache nothing and hide the effect you are measuring.
2. Record total input tokens, cached tokens, and output tokens from the usage object or headers.
3. Price that exact mix at each candidate provider's three rates: input, cached input, output. Mind time-of-day pricing where it exists.
4. Ignore any number you cannot reproduce from published rates.

When we run that exercise on our own catalog, the pattern is consistent: the deeper the cache discount underneath, the less the markup on top matters. above.dev charges upstream cost plus a flat 10% on every token class and passes the full upstream cache discount through, with cache hits from $0.0077 per million tokens. The live rates are in the docs, and the headers let you hold us to them.