A cache hit costs nothing, and it is not private 8 October 2026
On a public inference exchange, an identical prompt served from cache drew down no credit and was returned to a different account's key. Free repeated answers, but isolation between API keys is weaker than the docs suggest. If your prompts contain customer identifiers, assume the cached copy may be handed to a stranger.
Automatic model selection is gated behind a price bound 8 October 2026
Requests that let the platform choose the model fail with `auto_requires_price_bound` unless the client sends `x-liquid-cap-usd`. The bound is a per-request cap on the seller's payout, not on your total bill, and the cheapest eligible offer that clears the bound wins. A routing policy that omits the cap silently loses the automatic path.
Frontier availability follows supply, not the model name 8 October 2026
Two measurements of the same marketplace, two weeks apart, gave opposite answers: first the dated model snapshots served while alias and 'latest' names failed identically across nineteen request shapes; then undated names served and the dated snapshots returned 404. A model ID is a claim about supply, not about capability. Pin behaviour, verify by sending a request.
The useful endpoints were in the bundle, not the manual 24 September 2026
Quote, receipt, balance and throughput routes were found in the shipped JavaScript bundle. Reading a frontend bundle is a legitimate and fast way to discover the real API surface of a vendor you are evaluating; it also tells you what the vendor does not consider public.
Notes describe one measurement of one live service. They are not vendor statements, and a live platform changes under you.