Question 1

What does “cache hit” mean?

Accepted Answer

Think of each request’s full input as a book: - First request: the provider reads the whole book at the normal input rate and remembers the early chapters on its servers. - Later requests: if the prefix is byte-identical, the provider only fully processes new pages; the reused prefix is billed at a lower cached input rate. A cache hit = the provider reuses work from a previous identical prefix. SupaNexus does not store your chat history or run Prompt Cache. Caching and hit detection are handled by the model provider. SupaNexus reads usage from the response and bills using the platform’s three-tier prices. Related: Billing & balance · Chat Completions — usage

Question 2

Where do I see prices? What are the three tiers?

Accepted Answer

When operators configure a cached input price, Models → model detail → Pricing shows: If cached input pricing is unset or zero, tiered pricing is disabled — all input is billed at the normal input rate (not free). OpenAPI `GET /v1/models` does not include sell prices; use the Models marketplace — see Model pricing. Related: Model pricing

Question 3

How does SupaNexus bill cache hits?

Accepted Answer

Per request (simplified): SupaNexus reads cached counts from the response `usage`, for example: - OpenAI: `usage.prompt_tokens_details.cached_tokens` - Anthropic: `usage.cache_read_input_tokens` Settlement still uses a single charge amount; usage details may include `cached_input_tokens` when enabled. Related: Usage, bills & alerts

Question 4

How can my app actually hit cache and save money?

Accepted Answer

SupaNexus meters and bills; your app controls hit rate: 1. Multi-turn chat: keep system prompts, long docs, and history unchanged; only append new messages. 2. Stable prefix first: fixed system prompts and RAG context at the start of `messages`. 3. Stable user id: use a consistent OpenAI `user` field to improve cache hit rates. 4. Anthropic: add `cache_control` on blocks you want cached — see Messages. The prefix must match byte-for-byte (even an extra space can miss).

Question 5

Does this affect 402 or 429?

Accepted Answer

- 402: insufficient available balance — unrelated to cache hits. - 429: spend cap or rate limit. Cache hits only change unit price for input tokens; they do not bypass balance or quota checks. Related: 402 vs 429

Cached Input & Prompt Cache Pricing