- Which models are counted exactly?
- OpenAI GPT and o-series models mapped to o200k_base or cl100k_base use js-tiktoken in the browser for exact counts. Claude, Gemini, Llama, Mistral, and DeepSeek are local estimates only.
- Is my prompt uploaded?
- No. Token counting runs entirely in your browser. We do not call Anthropic, Google, or OpenAI token-count APIs for this tool. Drafts stay in localStorage on your device.
- How accurate are Claude / Gemini estimates?
- They are local approximations (o200k-based proxy in the browser), not official vendor tokenizers. Treat them as planning aids. For billing-accurate counts, use each vendor’s API usage / count_tokens endpoint on your own infrastructure.
- How are costs estimated?
- We apply published list prices (USD per 1M input/output tokens). Prices as of 2026-03-01. Cached tokens, enterprise discounts, and region pricing can change your real bill.
- What is Chat mode?
- Chat mode counts System / User / Assistant messages with OpenAI-style ChatML framing overhead. Use it when budgeting multi-turn API calls. Non-OpenAI models still show Estimate.
- How do I split a long document?
- Open Split by token limit under Utilities. Set a max tokens per chunk; cuts prefer newlines when they fit inside the budget.
- Can I count many files at once?
- Yes — Batch Token Counter lets you multi-select local .md / .txt files, see a summary table, and export CSV. Files stay on your device.
- Can I set custom API prices or encodings?
- In Options, enable Custom USD / 1M for enterprise discounts, or force o200k_base / cl100k_base / r50k_base for tiktoken debugging. Preferences persist in localStorage.
- Does the PWA work offline?
- After the first visit (or install), the service worker precaches the app and tokenizer chunks so OpenAI-style counting still works offline. Open the site once online to warm the cache.
- Where is the blog?
- ThinkPHP serves /blog, /api/blog/*, and /backend. In local Nuxt dev these paths can be proxied to PHP.
