Skip to main content

obol_core/pricing/
mod.rs

1//! Price tables and the per-message cost kernel.
2
3pub mod as_of;
4pub mod refresh;
5pub mod store;
6pub use store::{current_path, embedded, pricing_dir, PriceStore};
7
8use crate::model::MessageUsage;
9use serde::{Deserialize, Serialize};
10
11/// Per-million-USD rates for one model. Tier-`above` fields apply when a single
12/// request's input exceeds `tier_boundary`.
13#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
14pub struct ModelPrice {
15    pub input: f64,
16    pub output: f64,
17    #[serde(default)]
18    pub cache_read: f64,
19    #[serde(default)]
20    pub cache_write: f64, // 5-minute cache write
21    #[serde(default)]
22    pub cache_write_1h: Option<f64>,
23    #[serde(default)]
24    pub tier_boundary: Option<u64>,
25    #[serde(default)]
26    pub input_above: Option<f64>,
27    #[serde(default)]
28    pub output_above: Option<f64>,
29    #[serde(default)]
30    pub cache_read_above: Option<f64>,
31    #[serde(default)]
32    pub cache_write_above: Option<f64>,
33}
34
35/// USD cost of one message's tokens under `price`.
36pub fn cost_for(price: &ModelPrice, u: &MessageUsage) -> f64 {
37    let above = price
38        .tier_boundary
39        .is_some_and(|b| u.request_input_tokens > b);
40    let pick = |base: f64, over: Option<f64>| if above { over.unwrap_or(base) } else { base };
41
42    let r_in = pick(price.input, price.input_above);
43    let r_out = pick(price.output, price.output_above);
44    let r_cr = pick(price.cache_read, price.cache_read_above);
45    let r_cw5 = pick(price.cache_write, price.cache_write_above);
46    let r_cw1 = price.cache_write_1h.unwrap_or(price.cache_write);
47
48    (u.input_uncached as f64 * r_in
49        + u.output as f64 * r_out
50        + u.cache_read as f64 * r_cr
51        + u.cache_write_5m as f64 * r_cw5
52        + u.cache_write_1h as f64 * r_cw1)
53        / 1_000_000.0
54}
55
56#[cfg(test)]
57mod tests {
58    use super::*;
59    use crate::model::Provider;
60
61    fn usage(input: u64, cr: u64, cw5: u64, out: u64, req_in: u64) -> MessageUsage {
62        MessageUsage {
63            model: "claude-opus-4-8".into(),
64            provider: Provider::Anthropic,
65            namespace: "litellm".into(),
66            input_uncached: input,
67            cache_read: cr,
68            cache_write_5m: cw5,
69            cache_write_1h: 0,
70            output: out,
71            request_input_tokens: req_in,
72            service_tier: Some("standard".into()),
73        }
74    }
75
76    // claude-opus-4-8: input 5, output 25, cache_read 0.5, cache_write 6.25 per-million.
77    fn opus() -> ModelPrice {
78        ModelPrice {
79            input: 5.0,
80            output: 25.0,
81            cache_read: 0.5,
82            cache_write: 6.25,
83            cache_write_1h: Some(10.0),
84            tier_boundary: None,
85            input_above: None,
86            output_above: None,
87            cache_read_above: None,
88            cache_write_above: None,
89        }
90    }
91
92    #[test]
93    fn flat_model_costs_each_bucket() {
94        // 1M uncached in, 1M cache_read, 1M cache_write_5m, 1M out
95        let u = usage(1_000_000, 1_000_000, 1_000_000, 1_000_000, 3_000_000);
96        let c = cost_for(&opus(), &u);
97        assert!((c - (5.0 + 0.5 + 6.25 + 25.0)).abs() < 1e-9, "got {c}");
98    }
99
100    #[test]
101    fn tiered_model_switches_rates_above_boundary() {
102        // sonnet-4-5: base input 3, above-200k input 6, output 15/22.5
103        let price = ModelPrice {
104            input: 3.0,
105            output: 15.0,
106            cache_read: 0.3,
107            cache_write: 3.75,
108            cache_write_1h: None,
109            tier_boundary: Some(200_000),
110            input_above: Some(6.0),
111            output_above: Some(22.5),
112            cache_read_above: Some(0.6),
113            cache_write_above: Some(7.5),
114        };
115        // request input 300k > 200k -> above rates
116        let u = usage(1_000_000, 0, 0, 1_000_000, 300_000);
117        let c = cost_for(&price, &u);
118        assert!((c - (6.0 + 22.5)).abs() < 1e-9, "got {c}");
119        // request input 100k < 200k -> base rates
120        let u2 = usage(1_000_000, 0, 0, 1_000_000, 100_000);
121        let c2 = cost_for(&price, &u2);
122        assert!((c2 - (3.0 + 15.0)).abs() < 1e-9, "got {c2}");
123    }
124}