@jay0073/tokencalculator-core

The shared TypeScript calculation engine behind a privacy-first, browser-based token calculator. It provides local token measurement, model cost comparison, image-token formulas, context-window analysis, and workload forecasting for Node.js and browser projects.

Web GUI: https://tokencalculator.dev

The web GUI also counts combined workloads from pasted text, readable PDF and DOCX documents, source/data files, and supported images. Document extraction and image-dimension processing stay in the browser.

npm install @jay0073/tokencalculator-core

import { compareModels, measureModel, projectWorkload } from '@jay0073/tokencalculator-core';

const measurement = await measureModel(
  { text: 'A prompt or extracted document body', outputTokens: 2_000 },
  'gpt-5.6-terra',
);

const comparison = await compareModels(
  { text: 'The same workload', outputTokens: 2_000 },
  ['gpt-5.6-terra', 'claude-sonnet-5', 'gemini-3.1-flash-lite'],
);

const multiFileComparison = await compareModels(
  {
    files: [
      { text: 'Text extracted from the first uploaded file' },
      { text: 'Text extracted from the second uploaded file' },
      { image: { width: 1600, height: 900, detail: 'high' } },
    ],
    outputTokens: 2_000,
  },
  ['gpt-5.6-terra', 'claude-sonnet-5'],
);

const monthly = projectWorkload(measurement.cost.total, 1_000);

OpenAI-compatible text uses exact local o200k_base BPE tokenization through js-tiktoken. Other tokenizer families use deterministic Provider-Calibrated UTF-8 Token Projections, and images use provider-family formulas. Low-level helpers such as calculateTokenCost, calculateContextUsage, and resolveModelRates are also exported.

The bundled catalog supplies convenient model defaults. Pricing and model limits change frequently, so production callers can override rates and should treat provider invoices as authoritative.

The core accepts any number of already-extracted text files and image dimensions through files; browser or workflow adapters remain responsible for reading the original file bytes. The older singular image input remains supported, and images is available when file metadata is unnecessary.

All calculations are local. No prompt, token count, or file content is transmitted.