Struct Registry 

pub struct Registry {
    pub models: HashMap<String, ModelPricing>,
}

Fields

models: HashMap<String, ModelPricing>

Implementations

impl Registry

pub fn load_embedded() -> Result<Self>

pub fn from_json_str(raw: &str) -> Result<Self>

pub fn get(&self, model: &str) -> Option<&ModelPricing>

pub fn estimate_cost( &self, model: &str, input_tokens: u32, output_tokens: u32, ) -> Option<f64>

pub fn estimate_cost_with_cache( &self, model: &str, input_tokens: u32, output_tokens: u32, cache_read_tokens: u32, cache_creation_tokens: u32, ) -> Option<f64>

Estimates cost including prompt-cache read and cache-creation tokens.

Cache read tokens are priced at cache_read_cost_per_1k and cache creation tokens at cache_creation_cost_per_1k when the model publishes that pricing. When a model has no cache-specific pricing, cache tokens fall back to being billed at the model’s normal input rate (input_cost_per_1k) rather than being treated as free, since providers bill uncached-equivalent tokens at the input rate whenever they don’t publish a discounted cache rate.

Cache-creation tokens are priced at the base cache-write rate, which is the 5-minute TTL rate. Writes with a 1-hour TTL and long-context tier rates (for example *_above_200k_tokens prices) are not modeled, so the estimate is low for those requests.

Token accounting contract

input_tokens, cache_read_tokens, and cache_creation_tokens must be disjoint counts that sum to the total prompt tokens billed — input_tokens must exclude any tokens already counted in cache_read_tokens or cache_creation_tokens.

Providers report prompt tokens in different shapes, and the crate’s Usage prompt_tokens keeps each provider’s shape as reported:

• Anthropic: input_tokens already excludes cache_read_input_tokens and cache_creation_input_tokens, which are separate fields. Pass these counts directly.
• OpenAI-compatible: prompt_tokens (or input_tokens on the Responses API) includes cached tokens; cached_tokens in the details object is a subset of it, not an additive count.
• Gemini: promptTokenCount includes cachedContentTokenCount.

For OpenAI-compatible and Gemini usage, subtract the cached count from the prompt count before you pass it as input_tokens. If you do not, the cached tokens are billed twice: once at the input rate and once at the cache rate.

Trait Implementations

impl Clone for Registry

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for Registry

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.