Understand provider caching and model prices
Administrators inspect model prices in the Account AI catalog. Model administration permissions apply. See the shared catalog before editing prices.
Complexity and the backend sorter apply provider-specific cache options. There is no application answer cache. An admitted inference still makes a provider request. Only provider usage proves a cache hit. Prompt length or an earlier request does not prove one.
| Provider | Implemented behavior |
|---|---|
| Native OpenAI | GPT-5.6 and later Responses models receive implicit mode and 30m TTL. Older models use provider defaults. A compatible endpoint does not establish native capability. |
| Anthropic | Complexity uses prompt-cache middleware with 5m TTL. Sorter adds a 5m breakpoint to its stable system prefix. |
| Implicit caching. No explicit cache resource, upload, TTL refresh, or storage operation. | |
| OpenRouter | Anthropic model IDs receive request-level 5m cache_control. Other models use router/provider implicit behavior. |
| Other compatible endpoint | No foreign cache options. Accounting can still include reported cache usage. |
These are application request rules, not a guarantee of a provider cache hit. Tool order and the system prefix remain stable. The application adds no padding merely to reach a cache minimum. Changing selected tools can change the prefix.
Usage and prices
Input totals include cache read/write tokens. Output totals include reasoning. Ordinary input subtracts cache read, base write, and one-hour write from total input. Reasoning is an output subset, not a second charge. Normalization adds cache categories to Anthropic's uncached input count.
The model editor stores USD per million tokens for input, output, cache read, cache write, and one-hour cache write. Storage uses USD per million tokens per hour. Gemini implicit caching creates no explicit storage operation. Blank means unknown. Zero means explicitly free. Missing usage is not zero tokens.
An admitted run retains provider/model identity, adapter type, and fallback prices. Each actual request records current model prices while that model exists. After catalog deletion, admission prices remain available. Changing provider type does not reinterpret a running adapter's usage.
For long context, the highest strictly exceeded input threshold applies to the entire request, including cache input. Input, output, and cache multipliers are separate. Administrator prices and tiers take precedence. No background price synchronization runs.
Versioned catalog seeds contain standard text cache tariffs for retained native OpenAI GPT-6 Sol/Luna, Claude, and Gemini 2.5 models. They create no new model and preserve manual prices, including manually unknown prices. OpenRouter discovery converts published input_cache_read and input_cache_write prices to per-million units. Unpublished prices remain unknown.
Custom contracts, regional, priority, and audio tariffs need administrator configuration. A stored price snapshot is not a provider invoice. Explicit Gemini caching is unavailable in this implementation.