token-cost-time

token-cost-time

token-cost-time is a lightweight, dependency-free estimator for LLM task token usage, cost, and duration.

It uses rule-based task classification plus baseline priors and optional local profile data.

Features

• Rule-based task classification (no LLM calls)
• Cost + duration estimation by model
• Confidence ramp from cold-start to learned profile data
• Quality confidence adjusted by retry behavior
• Local execution recording for continuous calibration

Install / Run

From this module directory:

npm test
node cli/index.js calibrate --objective "summarize this document" --model claude-haiku-3
node cli/index.js compare --objective "implement a React component" --models claude-haiku-3,claude-sonnet-4,gpt-4o
node cli/index.js record --model claude-haiku-3 --task-class summarization --input-tokens 1340 --output-tokens 290 --duration-ms 4100 --cost 0.00071

API

• calibrate(objective, model, profilePath = null)
• record(executionData, profilePath = null)
• compare(objective, models, profilePath = null, sortBy = 'cost')

Library usage example:

import { calibrate } from './src/calibrate.js';

const result = calibrate('summarize this architecture memo', 'claude-haiku-3');
// result.taskClass => 'summarization'
// result.estimate.costUsd => estimated dollar cost
// result.confidence => 0.2 cold start, rises to 0.95 with 50+ recorded runs

If profilePath is omitted, profile defaults to:

• ~/.token-cost-time/profile.json

Execution log defaults to:

• ~/.token-cost-time/execution-log.jsonl

Task Classes

• summarization
• code_generation
• code_audit
• reasoning
• creative
• extraction
• conversation (fallback)

Notes

• ESM modules throughout.
• Node built-ins only.
• File paths are resolved with os.homedir() for cross-platform compatibility.