Local math · your prompt never leaves the browser
LLM Token Cost Calculator
Paste a prompt — or type a token count like 2500 — and see the estimated input and output tokens plus what a single call costs on GPT, Claude, Gemini, DeepSeek and Llama, sorted cheapest first.
Estimate one call
Everything below is computed in this tab — paste freely, no key, no upload, results in milliseconds.
Default 1,000-token example — the numbers above update as you type.
Snapshot prices are loaded; the table sorts by total cost, cheapest first.
Price snapshot: 2026-09-01 · Rows are sorted by total cost, cheapest first · These rates move every few months, so treat them as editable starting points, not a billing source.
Edit prices, add a model, or reset to the snapshot
How the estimate is built
Three steps, all client-side, all instant.
Your text is bucketed by character class: ASCII letters, digits and spaces run about 4 characters per token, CJK about 1.5, code punctuation and symbols around 2. Emoji and other scripts get their own bucket. A bare number such as 2500 or 2k skips estimation entirely.
The ratio you pick (0.25× to 5×) multiplies the input count to project output tokens — the part of the bill that usually dominates, because output rates are several times input rates on most providers.
For each selected model: input/1M × $in + output/1M × $out, plus the same figure ×1,000 for batch runs. Rows are sorted by total, so the cheapest option is always on top.
Why developers use it instead of a pricing page
There is no backend call and no key: the page loads once and does all arithmetic in your tab, so confidential system prompts and customer data stay on your machine.
Vendors change rates constantly and fresh models carry placeholder numbers. Every rate is editable, custom models can be added, and your edits are remembered in this browser.
Flip the output ratio from 0.25× to 5× and watch the order change — chatty models with expensive output tokens lose their edge on generation-heavy workloads.
Each row shows the per-call price and the price for 1,000 calls, which is the number you actually put in a spreadsheet when sizing a scrape or eval run.
Already know the exact count from tiktoken or your provider dashboard? Type it and the cost table becomes exact, with the heuristic out of the picture.
On narrow screens the comparison becomes one card per model with labelled values, and the token totals stay pinned at the top while you scroll.
FAQ
Is my prompt uploaded or stored anywhere?
No. The page makes no network requests, has no accounts and no analytics on your text. Nothing you paste leaves the browser tab, and closing it discards everything.
How accurate is the token estimate?
Treat it as roughly ±10% on ordinary prose. Code, JSON, emoji, and non-English text drift more because real tokenizers differ. For an exact number, count tokens with the provider's own tokenizer and type the result into the box — the cost math then matches to the cent.
Why do some models show "price unverified"?
Newly released models often have promotional or unannounced rates, so those rows carry placeholder values from the snapshot date. Open the price panel, correct the rate, and the badge disappears in favour of your number.
Do you include batch, cache or context-length discounts?
No — this is the list price for a standard single call, which is the right baseline for comparing models. Batch discounts of roughly 50% or cached-input discounts are provider-specific and easy to apply on top of the per-call figure.
Can I use the numbers as a billing reference?
Use it for planning, not for invoices. Estimates and hand-edited prices drift from real billing, which also depends on retries, prompt caching, reasoning tokens and provider rounding. Always confirm against the official pricing page and your usage dashboard.
What happens if I paste something enormous?
Input is capped at 100,000 characters. Beyond that the page tells you it measured only the first 100,000 instead of quietly reporting a wrong total, and the token counter above makes the applied length obvious.