About
Why this exists
Counting tokens across multiple AI providers should not require five different tools. OpenAI publishes a tokenizer. Anthropic exposes one in their API. Google has another. Open-source models each ship their own. The result is that comparing cost across models, the single most common AI cost question, is needlessly hard.
This site fixes that. Paste once, see counts and cost across every major model, with honest accuracy labels per row.
What "honest accuracy labels" means
Most multi-provider token counters quietly use one tokenizer for everything, which produces counts that are wrong by 5-20% for any model that isn't the one the tokenizer belongs to. A few wave their hands with a single site-wide accuracy claim. We do neither.
Every row in the results table carries its own confidence label. OpenAI counts run in your browser through the official o200k_base tokenizer via js-tiktoken: exact. Anthropic and Gemini counts come from each provider's official count_tokens endpoint through a serverless proxy: exact, identical to what billing uses. Llama counts use real BPE via llama-tokenizer-js: exact. Mistral, DeepSeek, Qwen, and GLM use a character-class-aware heuristic labeled ≈±3%, because no clean JavaScript implementation of their tokenizers exists yet. When that changes, the labels will change, and the pricing changelog will record it.
How pricing stays current
Prices are hand-snapshotted from each provider's published rate card and tracked in a git-versioned data file. Every change is dated and sourced in the public pricing changelog, including our own mistakes: when we shipped Claude Opus with the prior generation's pricing, the correction went in the log where anyone can see it. The "Pricing as of" badge in the counter tells you exactly how fresh the numbers are. If a price looks stale, the changelog tells you whether it is.
Who built it
Clinton Patrick. Senior Data Governance Engineer and indie web app builder. This site is part of a small portfolio of free, fast, no-signup utilities:
- whereisdns.com, global DNS propagation checker
- readabilitycheck.com, readability scoring
- epochcalc.com, epoch and time conversion toolkit
The common thread: each tool does one job, loads in under a second, and works without an account.
Stack
Vanilla HTML, CSS, and JavaScript. No framework. The whole site is served by a single Cloudflare Worker; tokenizer vocabularies load lazily as separate chunks so the initial page stays around 11 KB gzipped. No prompt content is ever logged, by us or by the proxy. Pricing data will be open-source in v1.1.
Contact
Find typos, broken counts, stale prices, or missing models? Email ideas@boringhaus.com. Pricing corrections are usually live within a day, and the changelog credits the catch.