# iron_cost - Specification
**Module:** iron_cost
**Layer:** 3 (Feature)
**Status:** Active
> **Specification Philosophy:** This specification focuses on architectural-level design and well-established knowledge. It describes what the module does and why, not implementation details or algorithms. Implementation constraints are minimal to allow flexibility. For detailed requirements, see spec/-archived_detailed_spec.md.
---
## Responsibility
LLM pricing data and cost calculation. Provides pricing data for multiple LLM providers, calculates costs based on token usage using integer arithmetic (microdollars) to avoid floating-point precision errors.
---
## Scope
**In Scope:**
- Multi-provider pricing data (LiteLLM pricing source)
- Cost calculation (tokens → microdollars, with USD conversion)
- Pre-reservation cost estimation for budget enforcement
- Per-model pricing with input/output token rates
- Currency conversion utilities (USD ↔ microdollars)
- Budget tracking with CostController (in-memory, thread-safe)
**Out of Scope:**
- Token counting (handled by LLM response)
- Multi-currency support (USD only)
- Cost forecasting (see analytics features)
- Budget persistence (uses in-memory tracking only)
---
## Dependencies
**Required Modules:**
- iron_types - Foundation types
**Required External:**
- arc_swap - Lock-free concurrent access to pricing data
- serde/serde_json - Pricing data serialization
**Optional:**
- None
---
## Core Concepts
**Key Components:**
- **PricingManager:** Thread-safe pricing data store with lock-free reads (ArcSwap)
- **Model:** Pricing info for single LLM model with cost calculation methods
- **Converter:** USD ↔ microdollars conversion utilities
- **Embedded Pricing:** LiteLLM pricing JSON embedded at compile time
- **CostController:** Thread-safe budget tracking with atomic operations
- **CostError:** Error types for budget enforcement (BudgetExceeded)
**Microdollar Precision:**
- 1 USD = 1,000,000 microdollars
- Per-token costs stored as "microdollars per million tokens" (u64)
- All cost calculations use integer arithmetic to avoid floating-point errors
- Example: $0.00000125/token = 1,250,000 micros/million tokens
**Cost Calculation:**
- Actual cost (micros): `(input_tokens * input_micros_per_M + output_tokens * output_micros_per_M) / 1,000,000`
- Max cost (pre-reservation): Same formula using max_output_tokens limit
- Default max_output_tokens: 128,000 (when model has no limit info)
**API Surface:**
- `calculate_cost_micros(input, output) -> u64` - Primary integer API
- `calculate_cost(input, output) -> f64` - Convenience USD wrapper
- `calculate_max_cost_micros(input, max_output) -> u64` - Budget pre-reservation
- `converter::usd_to_micros(f64) -> u64` - Currency conversion
- `converter::micros_to_usd(u64) -> f64` - Currency conversion
**CostController API:**
- `CostController::new(budget_usd: f64)` - Create with budget limit
- `check_budget() -> Result<(), CostError>` - Check if under limit
- `add_spend(cost_usd: f64)` - Add cost after request
- `add_spend_micros(cost_micros: u64)` - Add cost in microdollars
- `set_budget(budget_usd: f64)` - Update budget limit at runtime
- `total_spent() -> f64` - Get total spent in USD
- `budget_limit() -> f64` - Get current limit in USD
- `get_status() -> (f64, f64)` - Get (spent, limit) tuple
---
## Integration Points
**Used by:**
- iron_runtime/LlmRouter - Uses CostController for budget enforcement, PricingManager for cost calculation
- iron_control_api - Cost reporting endpoints
**Uses:**
- Embedded LiteLLM pricing data (asset/pricing.json)
---
## Testing
**Unit Tests:** `tests/budget_test.rs`
- CostController creation and initialization
- Budget checking (under/at/over limit)
- Spending accumulation
- Budget modification at runtime
- Edge cases (zero budget, large amounts, microdollar precision)
---
*For detailed pricing and algorithms, see spec/-archived_detailed_spec.md*
*For budget protocol, see docs/protocol/005_budget_control_protocol.md*