modelcost

Calculate LLM API call costs from token usage using price catalogs from multiple sources.

Supported pricing sources:

• litellm (default)
• openrouter
• tokencost

Install

python -m pip install modelcost

CLI

The default command calculates cost, so you can omit the cost subcommand.

# Default (cost)
modelcost gpt-4o 1000 500

# Explicit cost (optional)
modelcost cost gpt-4o 1000 500

# All sources in one run
modelcost --source all gpt-4o 1000 500

# JSON output
modelcost --json gpt-4o 1000 500

Cached and reasoning tokens

Modern LLM APIs charge differently for cached input and reasoning output tokens. Pass them as optional flags — they default to 0 so existing usage is unchanged.

# Cached input tokens (served from prompt cache)
modelcost gpt-4o 1000 500 --cached-input-tokens 200

# Cache creation tokens (first-time cache writes)
modelcost gpt-4o 1000 500 --cache-creation-input-tokens 100

# Reasoning tokens (subset of output_tokens, e.g. o1/R1 thinking)
modelcost deepseek/deepseek-r1 2000 5000 --reasoning-tokens 3000

# All together
modelcost gpt-4.1-mini 1000 500 \
  --cached-input-tokens 200 \
  --cache-creation-input-tokens 100 \
  --reasoning-tokens 150

List models

modelcost models
modelcost models --source openrouter
modelcost models --filter gpt
modelcost models --json

Help

modelcost --help
modelcost models --help

Library

from modelcost.calculator import calculate_cost, list_models

# Basic usage (backward compatible)
result = calculate_cost("gpt-4o", 1000, 500)

for source in result.available_sources:
    print(f"{source.source}: ${source.total_cost_usd:.6f}")

# With cached and reasoning tokens
result = calculate_cost(
    "gpt-4.1-mini",
    input_tokens=1000,
    output_tokens=500,
    cached_input_tokens=200,
    cache_creation_input_tokens=100,
    reasoning_tokens=150,
)

s = result.sources[0]
print(f"${s.total_cost_usd:.6f}")
print(f"  cache read:  ${s.price_per_million_cache_read}/M")
print(f"  reasoning:   ${s.price_per_million_reasoning}/M")

models = list_models("openrouter")

Output details

calculate_cost() returns a CostResult with:

• model, input_tokens, output_tokens
• cached_input_tokens, cache_creation_input_tokens, reasoning_tokens (0 when not used)
• sources: list of SourceCost objects
• available_sources: only sources with prices found

Each SourceCost includes:

• source
• total_cost_usd
• price_per_million_input, price_per_million_output
• price_per_million_cache_read, price_per_million_cache_creation, price_per_million_reasoning (present only when the source has specific pricing for these)
• error (when not available)

Cost formula

All subset tokens (cached_input_tokens, cache_creation_input_tokens, reasoning_tokens) are treated as subsets of their parent total and are clamped accordingly:

text_input  = input_tokens  - cached_input - cache_creation
text_output = output_tokens - reasoning_tokens

total = text_input     * input_rate
      + cached_input   * cache_read_rate    (fallback: input_rate)
      + cache_creation * cache_creation_rate (fallback: input_rate)
      + text_output    * output_rate
      + reasoning      * reasoning_rate      (fallback: output_rate)

This matches how most APIs report usage — input_tokens and output_tokens are the totals including cached/reasoning, and the detail fields are subsets. When a specific rate is missing, the base rate for that category is used as fallback.

JSON output

--json / result.to_dict() includes the new fields only when non-zero:

{
  "model": "gpt-4.1-mini",
  "input_tokens": 1000,
  "output_tokens": 500,
  "cached_input_tokens": 200,
  "reasoning_tokens": 150,
  "costs": [
    {
      "source": "litellm",
      "total_cost_usd": 0.001148,
      "price_per_million_input": 0.4,
      "price_per_million_output": 1.6,
      "price_per_million_cache_read": 0.1,
      "price_per_million_cache_creation": 0.48,
      "price_per_million_reasoning": 1.6,
      "error": null
    }
  ]
}

Caching

openrouter responses are cached in ~/.modelcost_cache.json for 1 hour.

Notes

• Prices are fetched at runtime from the upstream catalogs.
• If a model is missing in a source, that source is marked as unavailable.
• Network sources are fetched in parallel for the all option.
• tokencost does not expose cache/reasoning pricing — when used with source="all", its cost may be higher than litellm for calls that include cached or reasoning tokens.