ironflow-core 4.3.0

//! Unified LLM cost attribution per step and workflow.
//!
//! Provides a provider-agnostic pricing interface ([`PricingSource`]) with a
//! built-in static implementation ([`StaticPricing`]) that covers all supported
//! model families. Cost is computed into a [`CostBreakdown`] with uncached
//! prompt, cache read, cache write and completion components, rounded to 6
//! decimal places. Per-model rates, including prompt-cache rates, are exposed
//! through [`ModelPricing`].
//!
//! The [`spawn_log`] helper emits cost telemetry in a fire-and-forget task so
//! tracking never blocks step execution.
//!
//! # Examples
//!
//! ```no_run
//! use ironflow_core::pricing::{CostBreakdown, PricingSource, StaticPricing};
//!
//! let pricing = StaticPricing::new();
//! let breakdown = CostBreakdown::compute(&pricing, "claude-sonnet-4-6", 1000, 500);
//! println!("total: ${:.6}", breakdown.total_usd);
//! ```

use tracing::{info, warn};

/// A source of per-model token pricing.
///
/// Implementations return the cost per million tokens (input, output) for a
/// given model identifier. A `None` return means the model is not in the
/// catalog; callers should apply a fallback.
///
/// # Examples
///
/// ```no_run
/// use ironflow_core::pricing::{PricingSource, StaticPricing};
///
/// let pricing = StaticPricing::new();
/// if let Some((input, output)) = pricing.price_per_1m("claude-sonnet-4-6") {
///     println!("input: ${input}/Mtok, output: ${output}/Mtok");
/// }
/// ```
pub trait PricingSource: Send + Sync {
    /// Return `(input_per_1m_usd, output_per_1m_usd)` for the given model.
    ///
    /// Returns `None` when the model is not in the catalog.
    fn price_per_1m(&self, model: &str) -> Option<(f64, f64)>;

    /// Return the full [`ModelPricing`] (including prompt-cache rates) for the
    /// given model.
    ///
    /// The default implementation derives it from
    /// [`price_per_1m`](PricingSource::price_per_1m) with
    /// [`ModelPricing::without_cache`], so cache tokens are billed at the
    /// input rate. Returns `None` when the model is not in the catalog.
    ///
    /// # Examples
    ///
    /// ```no_run
    /// use ironflow_core::pricing::{PricingSource, StaticPricing};
    ///
    /// let pricing = StaticPricing::new();
    /// if let Some(p) = pricing.model_pricing("claude-sonnet-4-6") {
    ///     println!("cache read: ${}/Mtok", p.cache_read_per_1m);
    /// }
    /// ```
    fn model_pricing(&self, model: &str) -> Option<ModelPricing> {
        self.price_per_1m(model)
            .map(|(i, o)| ModelPricing::without_cache(i, o))
    }
}

/// Per-million-token rates for a single model, in USD.
///
/// Separates uncached input, output, prompt-cache reads and prompt-cache
/// writes, which providers bill at different rates.
///
/// # Examples
///
/// ```
/// use ironflow_core::pricing::ModelPricing;
///
/// let p = ModelPricing::without_cache(3.0, 15.0);
/// assert_eq!(p.cache_read_per_1m, 3.0);
/// assert_eq!(p.cache_write_per_1m, 3.0);
/// ```
#[derive(Debug, Clone, Copy, PartialEq)]
pub struct ModelPricing {
    /// Price per million uncached input tokens.
    pub input_per_1m: f64,
    /// Price per million output tokens.
    pub output_per_1m: f64,
    /// Price per million input tokens served from the prompt cache.
    pub cache_read_per_1m: f64,
    /// Price per million input tokens written to the prompt cache.
    pub cache_write_per_1m: f64,
}

impl ModelPricing {
    /// Build pricing for a model without a known cache rate.
    ///
    /// Both cache rates are set equal to `input_per_1m`, so costs are never
    /// underestimated.
    ///
    /// # Examples
    ///
    /// ```
    /// use ironflow_core::pricing::ModelPricing;
    ///
    /// let p = ModelPricing::without_cache(0.5, 1.5);
    /// assert_eq!(p.input_per_1m, 0.5);
    /// assert_eq!(p.output_per_1m, 1.5);
    /// assert_eq!(p.cache_read_per_1m, 0.5);
    /// ```
    #[must_use]
    pub fn without_cache(input_per_1m: f64, output_per_1m: f64) -> Self {
        Self {
            input_per_1m,
            output_per_1m,
            cache_read_per_1m: input_per_1m,
            cache_write_per_1m: input_per_1m,
        }
    }
}

/// Computed cost for a single LLM call, split into uncached prompt, cache
/// read, cache write and completion.
///
/// All values are in USD, rounded to 6 decimal places.
///
/// # Examples
///
/// ```no_run
/// use ironflow_core::pricing::{CostBreakdown, StaticPricing};
///
/// let pricing = StaticPricing::new();
/// let bd = CostBreakdown::compute(&pricing, "claude-sonnet-4-6", 10_000, 2_000);
/// assert!(bd.total_usd > 0.0);
/// assert!((bd.total_usd - bd.prompt_usd - bd.completion_usd).abs() < 1e-9);
/// ```
#[derive(Debug, Clone, Copy, PartialEq)]
pub struct CostBreakdown {
    /// Cost of uncached prompt (input) tokens in USD.
    pub prompt_usd: f64,
    /// Cost of input tokens served from the prompt cache in USD.
    pub cache_read_usd: f64,
    /// Cost of input tokens written to the prompt cache in USD.
    pub cache_write_usd: f64,
    /// Cost of completion (output) tokens in USD.
    pub completion_usd: f64,
    /// Total cost in USD: sum of `prompt_usd`, `cache_read_usd`,
    /// `cache_write_usd` and `completion_usd`.
    pub total_usd: f64,
}

/// Round to 6 decimal places.
fn round6(v: f64) -> f64 {
    (v * 1_000_000.0).round() / 1_000_000.0
}

/// Conservative fallback price per million tokens (input, output).
/// Uses the Claude Sonnet rate ($3/$15), shared across Sonnet 4.5/4.6.
const SONNET_FALLBACK: (f64, f64) = (3.0, 15.0);

impl CostBreakdown {
    /// Compute the cost for an LLM call using the given pricing source.
    ///
    /// When the model is unknown the fallback Sonnet price is used and a
    /// warning is logged. The cost is never zero for a non-zero token count.
    ///
    /// # Examples
    ///
    /// ```no_run
    /// use ironflow_core::pricing::{CostBreakdown, StaticPricing};
    ///
    /// let pricing = StaticPricing::new();
    /// let bd = CostBreakdown::compute(&pricing, "claude-opus-5", 5000, 1000);
    /// println!("${:.6}", bd.total_usd);
    /// ```
    pub fn compute(
        source: &dyn PricingSource,
        model: &str,
        input_tokens: u64,
        output_tokens: u64,
    ) -> Self {
        Self::compute_with_cache(source, model, input_tokens, 0, 0, output_tokens)
    }

    /// Compute the cost for an LLM call, pricing prompt-cache tokens separately.
    ///
    /// `input_tokens` is the uncached input only. Cache reads and writes are
    /// billed at the rates returned by [`PricingSource::model_pricing`]. When
    /// the model is unknown the fallback Sonnet price is used (cache rates
    /// equal to the input rate) and a warning is logged.
    ///
    /// # Examples
    ///
    /// ```no_run
    /// use ironflow_core::pricing::{CostBreakdown, StaticPricing};
    ///
    /// let pricing = StaticPricing::new();
    /// let bd = CostBreakdown::compute_with_cache(
    ///     &pricing,
    ///     "claude-sonnet-4-6",
    ///     1_000,
    ///     50_000,
    ///     2_000,
    ///     500,
    /// );
    /// println!("cache read: ${:.6}", bd.cache_read_usd);
    /// ```
    pub fn compute_with_cache(
        source: &dyn PricingSource,
        model: &str,
        input_tokens: u64,
        cache_read_tokens: u64,
        cache_write_tokens: u64,
        output_tokens: u64,
    ) -> Self {
        let rates = source.model_pricing(model).unwrap_or_else(|| {
            warn!(
                model,
                fallback_input = SONNET_FALLBACK.0,
                fallback_output = SONNET_FALLBACK.1,
                "unknown model, falling back to Claude Sonnet pricing"
            );
            ModelPricing::without_cache(SONNET_FALLBACK.0, SONNET_FALLBACK.1)
        });

        let prompt_usd = round6(input_tokens as f64 / 1_000_000.0 * rates.input_per_1m);
        let cache_read_usd =
            round6(cache_read_tokens as f64 / 1_000_000.0 * rates.cache_read_per_1m);
        let cache_write_usd =
            round6(cache_write_tokens as f64 / 1_000_000.0 * rates.cache_write_per_1m);
        let completion_usd = round6(output_tokens as f64 / 1_000_000.0 * rates.output_per_1m);
        let total_usd = round6(prompt_usd + cache_read_usd + cache_write_usd + completion_usd);

        Self {
            prompt_usd,
            cache_read_usd,
            cache_write_usd,
            completion_usd,
            total_usd,
        }
    }
}

/// Static, hardcoded pricing table for all supported model families.
///
/// Resolution is by substring: the most specific (longest) matching entry
/// wins. For example, `"claude-sonnet-4-6[1m]"` matches the `"claude-sonnet-4-6"`
/// entry because the full model string contains that substring.
///
/// # Examples
///
/// ```no_run
/// use ironflow_core::pricing::{PricingSource, StaticPricing};
///
/// let p = StaticPricing::new();
/// // Exact match
/// assert!(p.price_per_1m("claude-opus-5").is_some());
/// // Substring match (model with [1m] suffix)
/// assert!(p.price_per_1m("claude-opus-5[1m]").is_some());
/// ```
pub struct StaticPricing {
    entries: Vec<(&'static str, f64, f64)>,
}

impl StaticPricing {
    /// Create a new static pricing table with all known models.
    #[must_use]
    pub fn new() -> Self {
        let mut entries = vec![
            // ── Anthropic ─────────────────────────────────────
            ("claude-fable-5", 10.0, 50.0),
            ("claude-fable-5-1", 10.0, 50.0),
            ("claude-mythos-5", 10.0, 50.0),
            ("claude-mythos-5-1", 10.0, 50.0),
            ("claude-opus-5-5", 4.0, 20.0),
            ("claude-sonnet-5-5", 2.0, 10.0),
            ("claude-opus-5", 5.0, 25.0),
            ("claude-sonnet-5", 2.0, 10.0),
            ("claude-opus-4-8", 5.0, 25.0),
            ("claude-opus-4-7", 5.0, 25.0),
            ("claude-sonnet-4-6", 3.0, 15.0),
            ("claude-opus-4-6", 5.0, 25.0),
            ("claude-sonnet-4-5", 3.0, 15.0),
            ("claude-haiku-4-5", 1.0, 5.0),
            // Aliases used by ClaudeCodeProvider
            ("sonnet", 3.0, 15.0),
            ("opus", 5.0, 25.0),
            ("haiku", 1.0, 5.0),
            // ── OpenAI ────────────────────────────────────────
            ("gpt-5.5", 5.0, 30.0),
            ("gpt-5.4-mini", 0.75, 4.5),
            ("gpt-5.4-nano", 0.20, 1.25),
            ("gpt-5.4", 2.5, 15.0),
            ("gpt-4.1-mini", 0.40, 1.60),
            ("gpt-4.1-nano", 0.10, 0.40),
            ("gpt-4.1", 2.0, 8.0),
            ("gpt-4o-mini", 0.15, 0.60),
            ("gpt-4o", 2.5, 10.0),
            // ── Mistral ───────────────────────────────────────
            ("mistral-medium-3.5", 1.5, 7.5),
            ("mistral-large", 0.50, 1.50),
            ("mistral-small", 0.10, 0.30),
            ("mistral-medium", 1.0, 3.0),
            ("codestral", 0.30, 0.90),
            // ── Google Gemini ─────────────────────────────────
            ("gemini-3.5-flash", 0.15, 0.60),
            ("gemini-3.1-flash-lite", 0.05, 0.20),
            ("gemini-2.5-pro", 1.25, 10.0),
            ("gemini-2.5-flash", 0.15, 0.60),
            ("gemini-2.5-flash-lite", 0.05, 0.20),
            // ── NVIDIA NIM ────────────────────────────────────
            ("nemotron-nano-9b", 0.04, 0.16),
            ("nemotron-super-49b", 0.10, 0.40),
            ("nemotron-ultra-253b", 0.90, 0.90),
        ];
        // Sort by key length descending for longest-match-first resolution.
        entries.sort_by_key(|e| std::cmp::Reverse(e.0.len()));
        Self { entries }
    }
}

impl Default for StaticPricing {
    fn default() -> Self {
        Self::new()
    }
}

/// Return `(read_multiplier, write_multiplier)` applied to the input rate for
/// prompt-cache tokens, keyed by the matched pricing table entry.
///
/// Returns `None` for model families without a known cache rate.
fn cache_multipliers(key: &str) -> Option<(f64, f64)> {
    if key.starts_with("claude-") || matches!(key, "sonnet" | "opus" | "haiku") {
        Some((0.1, 1.25))
    } else if key.starts_with("gpt-5") {
        Some((0.1, 1.0))
    } else if key.starts_with("gpt-4.1") {
        Some((0.25, 1.0))
    } else if key.starts_with("gpt-4o") {
        Some((0.5, 1.0))
    } else if key.starts_with("gemini-") {
        Some((0.25, 1.0))
    } else {
        None
    }
}

impl StaticPricing {
    /// Find the most specific table entry matching `model`.
    fn find_entry(&self, model: &str) -> Option<&(&'static str, f64, f64)> {
        self.entries.iter().find(|(key, _, _)| model.contains(key))
    }
}

impl PricingSource for StaticPricing {
    fn price_per_1m(&self, model: &str) -> Option<(f64, f64)> {
        self.find_entry(model).map(|(_, inp, out)| (*inp, *out))
    }

    fn model_pricing(&self, model: &str) -> Option<ModelPricing> {
        let (key, input, output) = *self.find_entry(model)?;
        Some(match cache_multipliers(key) {
            Some((read_mult, write_mult)) => ModelPricing {
                input_per_1m: input,
                output_per_1m: output,
                cache_read_per_1m: input * read_mult,
                cache_write_per_1m: input * write_mult,
            },
            None => ModelPricing::without_cache(input, output),
        })
    }
}

/// Emit a cost log line in a fire-and-forget tokio task.
///
/// The spawned task logs the cost breakdown via [`tracing::info!`] and never
/// blocks the calling step. Errors in the task are silently absorbed.
///
/// # Examples
///
/// ```no_run
/// use ironflow_core::pricing::{CostBreakdown, StaticPricing, spawn_log};
///
/// let pricing = StaticPricing::new();
/// let bd = CostBreakdown::compute(&pricing, "claude-opus-5", 5000, 1000);
/// spawn_log("my-step", "claude-opus-5", bd);
/// ```
pub fn spawn_log(step_name: &str, model: &str, breakdown: CostBreakdown) {
    let step = step_name.to_string();
    let model = model.to_string();
    tokio::spawn(async move {
        info!(
            step = %step,
            model = %model,
            prompt_usd = breakdown.prompt_usd,
            cache_read_usd = breakdown.cache_read_usd,
            cache_write_usd = breakdown.cache_write_usd,
            completion_usd = breakdown.completion_usd,
            total_usd = breakdown.total_usd,
            "agent step cost"
        );
    });
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn known_model_returns_correct_price() {
        let pricing = StaticPricing::new();
        let price = pricing.price_per_1m("claude-opus-5");
        assert_eq!(price, Some((5.0, 25.0)));
    }

    #[test]
    fn sonnet_5_and_opus_5_5_have_current_pricing() {
        let pricing = StaticPricing::new();
        assert_eq!(pricing.price_per_1m("claude-sonnet-5"), Some((2.0, 10.0)));
        assert_eq!(pricing.price_per_1m("claude-opus-5-5"), Some((4.0, 20.0)));
        assert_eq!(
            pricing.price_per_1m("claude-mythos-5-1"),
            Some((10.0, 50.0))
        );
    }

    #[test]
    fn sonnet_5_5_has_current_pricing() {
        let pricing = StaticPricing::new();
        assert_eq!(pricing.price_per_1m("claude-sonnet-5-5"), Some((2.0, 10.0)));
        assert_eq!(
            pricing.price_per_1m("claude-sonnet-5-5[1m]"),
            Some((2.0, 10.0))
        );
    }

    #[test]
    fn substring_resolution_most_specific_wins() {
        let pricing = StaticPricing::new();
        // "claude-sonnet-4-6[1m]" contains "claude-sonnet-4-6" (17 chars)
        // and also "sonnet" (6 chars). The longer key wins.
        let price = pricing.price_per_1m("claude-sonnet-4-6[1m]");
        assert_eq!(price, Some((3.0, 15.0)));

        // "gpt-4.1-mini" contains "gpt-4.1-mini" (12 chars) and "gpt-4.1" (7 chars).
        // The longer key wins.
        let price = pricing.price_per_1m("gpt-4.1-mini");
        assert_eq!(price, Some((0.40, 1.60)));
    }

    #[test]
    fn unknown_model_falls_back_to_sonnet() {
        let pricing = StaticPricing::new();
        // Unknown model -> None from price_per_1m
        assert!(pricing.price_per_1m("totally-unknown-model-xyz").is_none());

        // CostBreakdown::compute applies the fallback
        let bd =
            CostBreakdown::compute(&pricing, "totally-unknown-model-xyz", 1_000_000, 1_000_000);
        // Sonnet fallback: $3/Mtok input, $15/Mtok output
        assert_eq!(bd.prompt_usd, 3.0);
        assert_eq!(bd.completion_usd, 15.0);
        assert_eq!(bd.total_usd, 18.0);
    }

    #[test]
    fn cost_breakdown_rounds_to_six_decimals() {
        let pricing = StaticPricing::new();
        // 7 input tokens at $5/Mtok = 0.000035 (exact)
        // 3 output tokens at $25/Mtok = 0.000075 (exact)
        let bd = CostBreakdown::compute(&pricing, "claude-opus-5", 7, 3);
        assert_eq!(bd.prompt_usd, 0.000035);
        assert_eq!(bd.completion_usd, 0.000075);
        assert_eq!(bd.total_usd, 0.00011);

        // 1 input token at $3/Mtok = 0.000003 (exact)
        // 1 output token at $15/Mtok = 0.000015 (exact)
        let bd = CostBreakdown::compute(&pricing, "claude-sonnet-4-6", 1, 1);
        assert_eq!(bd.prompt_usd, 0.000003);
        assert_eq!(bd.completion_usd, 0.000015);
        assert_eq!(bd.total_usd, 0.000018);

        // Test rounding: 3 input tokens at $1.25/Mtok = 0.00000375 -> 0.000004
        let bd = CostBreakdown::compute(&pricing, "gemini-2.5-pro", 3, 0);
        assert_eq!(bd.prompt_usd, 0.000004);
    }

    #[test]
    fn zero_tokens_returns_zero_cost() {
        let pricing = StaticPricing::new();
        let bd = CostBreakdown::compute(&pricing, "claude-opus-5", 0, 0);
        assert_eq!(bd.prompt_usd, 0.0);
        assert_eq!(bd.completion_usd, 0.0);
        assert_eq!(bd.total_usd, 0.0);
    }

    #[tokio::test]
    async fn spawn_log_does_not_block() {
        use std::time::Duration;

        tokio::time::timeout(Duration::from_secs(5), async {
            let pricing = StaticPricing::new();
            let bd = CostBreakdown::compute(&pricing, "claude-opus-5", 1000, 500);
            spawn_log("test-step", "claude-opus-5", bd);
            tokio::task::yield_now().await;
        })
        .await
        .expect("spawn_log timed out");
    }

    #[test]
    fn all_known_models_have_prices() {
        let pricing = StaticPricing::new();
        let models = [
            "claude-fable-5",
            "claude-fable-5-1",
            "claude-mythos-5",
            "claude-mythos-5-1",
            "claude-opus-5-5",
            "claude-sonnet-5-5",
            "claude-opus-5",
            "claude-sonnet-5",
            "claude-opus-4-8",
            "claude-opus-4-7",
            "claude-sonnet-4-6",
            "claude-opus-4-6",
            "claude-sonnet-4-5",
            "claude-haiku-4-5",
            "sonnet",
            "opus",
            "haiku",
            "gpt-5.5",
            "gpt-5.4",
            "gpt-4.1",
            "gpt-4o",
            "gpt-4o-mini",
            "mistral-large",
            "mistral-small",
            "codestral",
            "gemini-2.5-pro",
            "gemini-2.5-flash",
            "nemotron-nano-9b",
            "nemotron-super-49b",
            "nemotron-ultra-253b",
        ];
        for model in models {
            assert!(
                pricing.price_per_1m(model).is_some(),
                "missing price for {model}"
            );
        }
    }

    #[test]
    fn entries_sorted_by_length_descending() {
        let pricing = StaticPricing::new();
        for window in pricing.entries.windows(2) {
            assert!(
                window[0].0.len() >= window[1].0.len(),
                "entries not sorted: {:?} before {:?}",
                window[0].0,
                window[1].0
            );
        }
    }

    #[test]
    fn default_matches_new() {
        let a = StaticPricing::new();
        let b = StaticPricing::default();
        assert_eq!(a.entries.len(), b.entries.len());
        for (ea, eb) in a.entries.iter().zip(b.entries.iter()) {
            assert_eq!(ea.0, eb.0);
            assert_eq!(ea.1, eb.1);
            assert_eq!(ea.2, eb.2);
        }
    }

    #[test]
    fn cost_breakdown_with_large_token_counts() {
        let pricing = StaticPricing::new();
        // 1M input tokens at $5/Mtok = $5.0
        // 500K output tokens at $25/Mtok = $12.5
        let bd = CostBreakdown::compute(&pricing, "claude-opus-5", 1_000_000, 500_000);
        assert_eq!(bd.prompt_usd, 5.0);
        assert_eq!(bd.completion_usd, 12.5);
        assert_eq!(bd.total_usd, 17.5);
    }

    fn assert_close(actual: f64, expected: f64) {
        assert!(
            (actual - expected).abs() < 1e-9,
            "expected {expected}, got {actual}"
        );
    }

    #[test]
    fn pricing_cache_read_rate_is_ten_percent_for_claude() {
        let pricing = StaticPricing::new();
        let p = pricing.model_pricing("claude-sonnet-4-6").unwrap();
        assert_close(p.input_per_1m, 3.0);
        assert_close(p.output_per_1m, 15.0);
        assert_close(p.cache_read_per_1m, 0.3);
        assert_close(p.cache_write_per_1m, 3.75);

        let alias = pricing.model_pricing("sonnet").unwrap();
        assert_close(alias.cache_read_per_1m, 0.3);
    }

    #[test]
    fn pricing_unknown_cache_rate_defaults_to_input() {
        let pricing = StaticPricing::new();
        let p = pricing.model_pricing("mistral-large").unwrap();
        assert_close(p.cache_read_per_1m, 0.50);
        assert_close(p.cache_write_per_1m, 0.50);
    }

    #[test]
    fn pricing_cache_read_breakdown_total_includes_four_parts() {
        let pricing = StaticPricing::new();
        let bd = CostBreakdown::compute_with_cache(
            &pricing,
            "claude-opus-5",
            1_000_000,
            1_000_000,
            1_000_000,
            1_000_000,
        );
        assert_eq!(bd.prompt_usd, 5.0);
        assert_eq!(bd.cache_read_usd, 0.5);
        assert_eq!(bd.cache_write_usd, 6.25);
        assert_eq!(bd.completion_usd, 25.0);
        assert_eq!(bd.total_usd, 36.75);
    }

    #[test]
    fn pricing_compute_without_cache_matches_legacy() {
        let pricing = StaticPricing::new();
        let bd = CostBreakdown::compute(&pricing, "claude-opus-5", 1_000_000, 500_000);
        assert_eq!(bd.cache_read_usd, 0.0);
        assert_eq!(bd.cache_write_usd, 0.0);
        assert_eq!(
            bd,
            CostBreakdown::compute_with_cache(&pricing, "claude-opus-5", 1_000_000, 0, 0, 500_000)
        );
    }

    #[test]
    fn pricing_unknown_model_cache_fallback() {
        let pricing = StaticPricing::new();
        assert!(pricing.model_pricing("totally-unknown-model-xyz").is_none());
        let bd = CostBreakdown::compute_with_cache(
            &pricing,
            "totally-unknown-model-xyz",
            0,
            1_000_000,
            1_000_000,
            0,
        );
        // Sonnet fallback input rate applies to both cache components.
        assert_eq!(bd.cache_read_usd, 3.0);
        assert_eq!(bd.cache_write_usd, 3.0);
        assert_eq!(bd.total_usd, 6.0);
    }

    #[test]
    fn pricing_openai_cache_read_rate() {
        let pricing = StaticPricing::new();
        let p = pricing.model_pricing("gpt-4.1").unwrap();
        assert_close(p.cache_read_per_1m, 0.5);
        assert_close(p.cache_write_per_1m, 2.0);

        let gpt5 = pricing.model_pricing("gpt-5.4").unwrap();
        assert_close(gpt5.cache_read_per_1m, 0.25);

        let gpt4o = pricing.model_pricing("gpt-4o").unwrap();
        assert_close(gpt4o.cache_read_per_1m, 1.25);

        let gemini = pricing.model_pricing("gemini-2.5-pro").unwrap();
        assert_close(gemini.cache_read_per_1m, 0.3125);
    }

    #[test]
    fn pricing_default_model_pricing_impl_without_cache() {
        struct FlatPricing;

        impl PricingSource for FlatPricing {
            fn price_per_1m(&self, model: &str) -> Option<(f64, f64)> {
                (model == "flat").then_some((2.0, 8.0))
            }
        }

        let p = FlatPricing.model_pricing("flat").unwrap();
        assert_eq!(p, ModelPricing::without_cache(2.0, 8.0));
        assert_eq!(p.cache_read_per_1m, 2.0);
        assert_eq!(p.cache_write_per_1m, 2.0);
        assert!(FlatPricing.model_pricing("other").is_none());
    }

    #[test]
    fn sonnet_fallback_never_zero() {
        let pricing = StaticPricing::new();
        // Even 1 token should produce a non-zero cost via fallback
        let bd = CostBreakdown::compute(&pricing, "unknown-model", 1, 0);
        assert!(bd.total_usd > 0.0);
    }
}