openai-cost

Calculate OpenAI API call cost from a usage block. Cache-aware (cached input tokens), supports GPT-5, GPT-4.1, o3, o4 model families. Bring your own pricing override. Zero runtime dependencies.

Why

Every usage logger ships its own copy of this. They get the cache math wrong, or they hard-code yesterday's pricing, or they break when OpenAI adds a new model. This crate is the small, dated, tested version you can reuse and override where you need to.

Usage

use openai_cost::{Usage, default_pricing};

let pricing = default_pricing("gpt-5").unwrap();

// From a Chat Completions response (prompt_tokens INCLUDES cached):
let usage = Usage::from_chat_completions(
    /* prompt_tokens     = */ 1000,
    /* completion_tokens = */  500,
    /* cached_tokens     = */  300,
);
let cost_usd = pricing.cost_for(&usage);

Or from the Responses API (already split):

use openai_cost::Usage;
let usage = Usage::from_responses_api(700, 500, 300);

Dated snapshots

Pass any dated snapshot id; the lookup strips the -YYYY-MM-DD suffix back to the base alias.

use openai_cost::default_pricing;
assert!(default_pricing("gpt-5-2026-04-01").is_some());
assert!(default_pricing("gpt-4o-mini-2024-07-18").is_some());

BYO pricing

use openai_cost::{Pricing, Usage};

let custom = Pricing {
    input_per_mtok: 1.25,
    output_per_mtok: 5.0,
    cached_input_per_mtok: 0.125,
};
let _ = custom.cost_for(&Usage::default());

Pricing notes

All rates are USD per 1,000,000 tokens. Pricing is dated as of 2026-Q2. Verify against https://openai.com/api/pricing/ before billing.

Features

serde — derive Serialize/Deserialize on Usage so you can parse a raw API response directly.

[dependencies]
openai-cost = { version = "0.1", features = ["serde"] }

License

Licensed under either of MIT or Apache-2.0 at your option.

openai-cost 0.1.0