Prompt Token & Cost Estimator

Overview

Count tokens with the real o200k_base BPE tokenizer bundled locally, forecast round-trip cost across 27 models from a static September 2026 price snapshot, and model prompt-caching savings. Everything runs offline; no text ever leaves the page.

How It Works

Paste a prompt or upload a text or code file (.txt, .md, .json, .csv, .py, .ts, .yaml and similar) to see live token, character, word, and context-load counts. Pick a tokenizer family, set expected output tokens, and read the cost table per 1,000 and 100,000 calls.

Step-by-Step Usage Guide

1. Paste text or upload a text or code file; the token, character, word, and context-load counters update live.
2. Choose a tokenizer family from the dropdown, or use the category filter pills and model search to narrow the board.
3. Enter a value in the 'Expected Output Tokens' input or pick a 300, 800, 2000, or 8000 preset to forecast round-trip cost.
4. Enable the prompt-caching toggle, set the cache-hit slider between 10 and 95 percent, and read cost per call, per 1,000, and per 100,000 calls.

Technical Specifications & Standards

Token counts are exact BPE counts for OpenAI's o200k_base encoding, the GPT-4o-era tokenizer, via the gpt-tokenizer library. The Tokenizer Family dropdown applies a stated heuristic to that one count — about 1.02 times for Anthropic and 1.08 times for Google — and every model row in the cost table uses the same resulting token count. The hardcoded price board covers 27 models, carries its catalog date, and drives a cost table with per-call, per-1,000, and per-100,000 figures plus a 'Best Value' highlight. Custom models you define persist in localStorage.

Targeted Use Cases

• Budgeting a batch job: paste the prompt, set an output preset, and read the per-100,000-calls figure for the model you plan to use.
• Comparing providers: compare the cost table rows, switch the tokenizer-family dropdown, and use the 'Best Value' highlight to shortlist models for the same workload.
• Sizing a context: watch the context-load indicator while uploading a long .md document to see whether the prompt fits before spending anything.
• Estimating caching gains: toggle prompt caching and sweep the 10 to 95 percent cache-hit slider to see how a repeated system prompt changes cost.

Notes & Gotchas

• Only the OpenAI family setting gives the exact o200k_base count. The Anthropic and Google settings are heuristics by construction, not those providers' real tokenizers, and DeepSeek, xAI and Mistral rows use whichever count is selected.
• The price board is hardcoded and dated, so prices drift. Verify current rates against the provider before committing to a budget.
• There is no separate DeepSeek, xAI or Mistral tokenizer setting; pick the family closest to your model and keep a safety margin.
• Define custom models for unlisted pricing; they persist in localStorage so repeat sessions keep your figures. The tool makes no network calls at all.

Frequently Asked Questions