tokenary

Minimal Python library to calculate LLM API costs based on the LiteLLM model catalog.

The bundled pricing data (tokenary/_generated.py, data/model_prices.generated.json) is generated from the LiteLLM model prices catalog. You are not dependent on a new tokenary release to pick up pricing updates: the generator ships as part of the installed package, so you can re-run it yourself at any time (see Generate pricing artifacts) to refresh the data against the latest LiteLLM catalog.

Installation

pip install tokenary

Usage

Functional API

import tokenary
from tokenary import ModelName

result = tokenary.calculate(
    model=ModelName.AZURE_GPT_3_5_TURBO,
    input_tokens=1000,
    output_tokens=500,
)

print(result.total_cost)
print(result.model_dump())

Request object

from tokenary import ModelName, UsageCostRequest, calculate

request = UsageCostRequest(
    model=ModelName.AZURE_GPT_3_5_TURBO,
    input_tokens=2000,
    output_tokens=800,
    reasoning_tokens=200,
)

result = calculate(request)
print(result.model_dump())

Reasoning tokens (e.g. o1)

from tokenary import ModelName, calculate

result = calculate(
    model=ModelName.O1,
    input_tokens=500,
    output_tokens=200,
    reasoning_tokens=300,
)

print(f"Total: ${result.total_cost:.6f}")
print(f"  Reasoning: ${result.reasoning_cost:.6f}")

All supported parameters

Parameter	Type	Description
model	ModelName	Model identifier
input_tokens	int	Number of input tokens
output_tokens	int	Number of output tokens
reasoning_tokens	int	Reasoning tokens (e.g. o1)
audio_input_tokens	int	Audio input tokens
generated_images	int	Number of generated images
code_interpreter_sessions	int	Code interpreter sessions
file_search_calls	int	File search API calls
file_search_gb_days	float	File search storage (GB-days)
vector_store_gb_days	float	Vector store storage (GB-days)

The returned CostBreakdown object contains per-category costs (input_cost, output_cost, reasoning_cost, …) and a total_cost, all in USD.

Generate pricing artifacts

All pricing data bundled with tokenary is generated from the LiteLLM model prices catalog, not hand-maintained. This means you can re-generate it locally at any time, e.g. to pick up upstream LiteLLM pricing changes before a new tokenary release ships them.

This works with a plain pip install tokenary too — the generator is part of the installed package, so no repo checkout is required:

python -m tokenary.generator

or via the tokenary-generate console script installed alongside the package:

tokenary-generate

If you're working from a repo checkout, there's also a helper script:

./scripts/generate_catalog.sh