The first Grok 4.5 billing checks looked alarming: multiple models appeared to cost 13 to 27 times their published rates. The numbers were precise, repeatable, and wrong.

The culprit was not an exotic token rule. It was a shared API key whose organization-wide billing counter kept moving while the test was trying to measure a few cents.

Disclosure: The author works with OrcaRouter and used it for the first-party observations cited here. Model access does not imply provider affiliation or endorsement.

Explore Grok models on OrcaRouter.

Why a shared counter breaks a tiny experiment

The measurement reads an account billing counter before and after a request. If other traffic is using the same key, the difference includes everyone else’s spend. In the failed runs, the counter was climbing by roughly $0.16 every 30 seconds, enough to overwhelm a small text conversation.

A useful clue was that unrelated control models appeared to have the same inflation factor. That is not a plausible per-model pricing pattern; it is an instrument problem.

The control that fixed it

The repaired procedure uses an exclusive API key and aborts unless two idle counter reads, 20 seconds apart, are byte-identical. Each measurement then follows settle, request, settle: wait for the counter to stop, make one request, wait for it to stop again, and take the delta.

With that condition met, predicted token usage multiplied by published prices reproduced the actual bill for Grok 4.5, GPT-5.6 Terra, and Kimi K3 to within rounding. The Grok three-turn conversation was $0.012346 predicted and $0.012346 billed.

The cache trap beside it

Billing arithmetic also has to split fresh and cached input. In one Grok conversation, 6,272 of 6,764 input tokens were cache hits. Charging all input at the headline fresh-input price would have overestimated the measured bill by 1.76x.

A price calculation without a clear cache convention is not necessarily dishonest. It is simply incomplete, and it can be badly wrong for a multi-turn application.

The reusable takeaway

Before publishing an LLM cost number, prove the meter is idle, use an exclusive key, retain the returned usage object, and state whether cache input is separated. This is unglamorous work, but it is the line between a cost anecdote and a usable measurement.

Explore the Models

Compare current AI models and routes in the OrcaRouter model catalog.

This article is a time-stamped preview or case study, not a provider endorsement or a general model ranking.

Sources

• xAI model documentation, checked 2026-07-31, for the stated 200K-token pricing threshold and rate schedule.
• First-party Grok 4.5 long-context billing measurement, 2026-07-30/31; six measurements, exclusive key, cache-free prompts, reproducible with grok_tier_test.py.
• First-party multi-vendor billing verification, 2026-07-30, reproducible with verify_billing.py.