runcost-ai 0.1.13

Alpha cost ledger utility for LLM and agent API responses.

RunCost

RunCost is a small alpha utility for answering one question:

What did this LLM or agent API call cost, and why?

It turns provider responses, framework usage objects, or normalized usage into a componentized cost ledger with input, cached input, output, reasoning, tool units, discounts, price sources, and warnings.

Install

Install from package registries:

pip install runcost-ai
npm install runcost
go get github.com/adamallcock/runcost/packages/go/ledger

Source checkout development paths:

python3 -m pip install git+https://github.com/adamallcock/runcost.git
PKG_TGZ=$(npm pack ./packages/javascript/core --silent)
npm install "./$PKG_TGZ"

The Python distribution name is runcost-ai; the import package and CLI are runcost. The npm package is runcost. The Go package is github.com/adamallcock/runcost/packages/go/ledger.

Default Price Catalog

RunCost includes an optional bundled source-cache catalog generated from llm-prices, LiteLLM, OpenRouter, models.dev, and reviewed official snapshots for targeted provider pricing gaps and redirects. It is package data, not fixture data.

Python: default_price_cards() / default_source_cache()

JavaScript/TypeScript: defaultPriceCards() / defaultSourceCache()

Go: DefaultPriceCards() / DefaultSourceCache()

One-Minute Examples

Python:

from runcost import from_response

response = {
    "model": "gpt-4.1-mini-2025-04-14",
    "usage": {
        "input_tokens": 36,
        "input_tokens_details": {"cached_tokens": 6},
        "output_tokens": 87,
        "output_tokens_details": {"reasoning_tokens": 12},
    },
}

price_cards = [{
    "schema_version": "0.1",
    "id": "openai:gpt-4.1-mini:example",
    "provider": "openai",
    "surface": "openai.responses",
    "model": "gpt-4.1-mini",
    "aliases": ["gpt-4.1-mini-2025-04-14"],
    "components": [
        {"usage_component": "input_uncached_tokens", "unit": "token", "price": {"amount": "0.40", "currency": "USD", "per": "1000000"}},
        {"usage_component": "input_cache_read_tokens", "unit": "token", "price": {"amount": "0.10", "currency": "USD", "per": "1000000"}},
        {"usage_component": "output_text_tokens", "unit": "token", "price": {"amount": "1.60", "currency": "USD", "per": "1000000"}},
        {"usage_component": "output_reasoning_tokens", "unit": "token", "price": {"amount": "1.60", "currency": "USD", "per": "1000000"}},
    ],
    "source": {"name": "example"},
}]

ledger = from_response(
    response,
    provider="openai",
    surface="openai.responses",
    model="gpt-4.1-mini",
    price_cards=price_cards,
)

print(ledger["total"])
print(ledger["components"])
print(ledger["warnings"])

TypeScript:

import { fromResponse } from "runcost";

// Using the same response and priceCards shape as the Python example above.
const ledger = fromResponse(response, {
  provider: "openai",
  surface: "openai.responses",
  model: "gpt-4.1-mini",
  priceCards
});

console.log(ledger.total);
console.log(ledger.components);
console.log(ledger.warnings);

Go:

package main

import (
    "fmt"

    ledger "github.com/adamallcock/runcost/packages/go/ledger"
)

func main() {
    priceCards := []any{
        ledger.Object{
            "schema_version": "0.1",
            "id":             "openai:gpt-4.1-mini:example",
            "provider":       "openai",
            "surface":        "openai.responses",
            "model":          "gpt-4.1-mini",
            "aliases":        []any{"gpt-4.1-mini-2025-04-14"},
            "components": []any{
                ledger.Object{
                    "usage_component": "input_uncached_tokens",
                    "unit":            "token",
                    "price": ledger.Object{"amount": "0.40", "currency": "USD", "per": "1000000"},
                },
                ledger.Object{
                    "usage_component": "output_text_tokens",
                    "unit":            "token",
                    "price": ledger.Object{"amount": "1.60", "currency": "USD", "per": "1000000"},
                },
            },
            "source": ledger.Object{"name": "example"},
        },
    }

    cost := ledger.FromResponse(
        ledger.Object{
            "model": "gpt-4.1-mini-2025-04-14",
            "usage": ledger.Object{
                "input_tokens":  36,
                "output_tokens": 87,
            },
        },
        ledger.Object{
            "provider": "openai",
            "surface":  "openai.responses",
            "model":    "gpt-4.1-mini",
        },
        priceCards,
        nil,
    )

    fmt.Println(cost["total"])
}

Already have normalized usage? Use the deterministic calculator directly:

from runcost import calculate_cost

ledger = calculate_cost(
    usage_ledger={
        "schema_version": "0.1",
        "provider": "openai",
        "surface": "openai.responses",
        "model": {"requested": "gpt-4.1-mini"},
        "components": [
            {"name": "input_uncached_tokens", "quantity": "30", "unit": "token"},
            {"name": "output_text_tokens", "quantity": "75", "unit": "token"},
        ],
    },
    price_cards=price_cards,
)

Main APIs

Job	Python	JavaScript/TypeScript	Go
Price normalized usage	calculate_cost(...)	calculateCost(options)	CalculateCost(options)
Price a provider response	from_response(...)	fromResponse(response, options)	FromResponse(response, options, priceCards, discountPolicies)
Aggregate call ledgers	aggregate_cost_ledgers(...)	aggregateCostLedgers(options)	AggregateCostLedgers(...)
Use framework outputs	from_langsmith_run(...), track_langchain_costs(...), and more	fromVercelAISDKStreamFinish(...), createRunCostVercelOnFinish(...), and more	FromLangSmithRun(...), FromSemanticKernelTelemetry(...), and more
Load price sources	price_cards_from_json_file(...), price_cards_from_openrouter_models(...)	priceCardsFromJSONFile(...), priceCardsFromOpenRouterModels(...)	PriceCardsFromJSONFile(...), PriceCardsFromOpenRouterModels(...)
Use bundled default catalog	default_price_cards()	defaultPriceCards()	DefaultPriceCards()
Add custom prices	Pass price_cards	Pass priceCards	Pass price_cards in options
Apply discounts	Pass discount_policies	Pass discountPolicies	Pass discount_policies in options
Audit decisions	debug_trace=True	debugTrace: true	"debug_trace": true
Fail on ambiguity	mode="strict"	mode: "strict"	mode: "strict"
CLI checks	runcost price-cards, runcost fixture-check	N/A	N/A

Supported Inputs

Fixture-backed surfaces include OpenAI Responses and Chat Completions, Anthropic Messages, OpenRouter, Gemini and Vertex generateContent, AWS Bedrock Converse, Cohere Chat and Rerank, OpenAI-compatible providers such as Meta, Groq, xAI, Mistral, DeepSeek, Azure OpenAI, and Hugging Face Inference Providers, plus selected framework objects from LangChain, Vercel AI SDK, OpenAI Agents SDK, LlamaIndex, Haystack, LiteLLM, AutoGen/AG2, LangSmith, Semantic Kernel, and OpenRouter SDK paths.

See supported surfaces for the current matrix.

Custom Prices And Discounts

RunCost treats provider pricing as data. You can pass user price cards for private rates, exact aliases, service tiers, long-context prices, historical effective dates, tool units, or internal billing units.

discounts = [{
    "schema_version": "0.1",
    "id": "openai-contract-4pct",
    "match": {"provider": "openai"},
    "adjustment": {"type": "percentage_discount", "value": "4"},
}]

The returned ledger records selected price sources, applied discounts, and any warning that prevents the total from being fully explained.

Fixtures are behavioral conformance tests, not a complete model-price database. Use source adapters, reviewed source-cache snapshots, or the optional bundled default catalog for upstream catalog data; see price data strategy.

Python:

from runcost import DEFAULT_PRICE_SOURCE_PRIORITY, default_price_cards, from_response

ledger = from_response(
    response,
    provider="openai",
    surface="openai.responses",
    model="gpt-4.1-mini",
    price_cards=default_price_cards(),
    price_source_priority=DEFAULT_PRICE_SOURCE_PRIORITY,
)

TypeScript:

import { DEFAULT_PRICE_SOURCE_PRIORITY, defaultPriceCards, fromResponse } from "runcost";

const ledger = fromResponse(response, {
  provider: "openai",
  surface: "openai.responses",
  model: "gpt-4.1-mini",
  priceCards: defaultPriceCards(),
  priceSourcePriority: DEFAULT_PRICE_SOURCE_PRIORITY
});

Warnings

RunCost is designed to be boring. When it cannot confidently price something, it returns a structured warning such as unknown_model, component_unpriced, price_stale, stream_usage_missing, or provider_reported_cost_mismatch. Use strict mode in tests or reconciliation flows when warnings should fail.

CLI

The Python package installs a lightweight CLI:

runcost price-cards --source-type user-pricing --input prices.json
runcost fixture-check fixtures/my-case.json

Read Next

• Quickstart
• Package installation
• Migration from hand-written formulas
• API reference
• Supported surfaces
• Custom pricing and discounts
• Source adapters
• Price data strategy
• Aggregation and streaming
• Warnings and limitations
• Contributing
• Security
• Changelog

Status

RunCost is alpha software. The core behavior is fixture-backed across Python, JavaScript/TypeScript, and Go, and alpha packages are published to PyPI, npm, and Go module tags. Smoke costs may use sample price cards; use provider exports or dashboard reconciliation before treating a total as invoice-exact.