---
title: "Token counter: GPT-4o, GPT-5 and GPT-4 tokens of any text, and what an MCP tool list costs"
description: "Paste a prompt, a document or the tool list of an MCP server. Exact counts for GPT-4o, GPT-4.1, GPT-5 and GPT-4, and each tool of a list counted on its own."
canonical: "https://scalebrowser.net/check/token-counter"
last_modified: "2026-10-23T18:22:53.000Z"
---

> ## Documentation Index
> Fetch the complete documentation index at: https://scalebrowser.net/llms.txt
> Use this file to discover all available pages before exploring further.

one check

# Token counter

Paste a prompt, a document or the tool list of an MCP server and see how many tokens it is for GPT models, and for Claude where Anthropic's own counter is available. A tool list is split per tool, because an agent pays for it on every request.

How are the tokens counted?

GPT models use public encodings: o200k\_base for GPT-4o, GPT-4.1, GPT-5 and the o-series, cl100k\_base for GPT-4 and GPT-3.5. Both are counted exactly on our server. Claude uses its own tokenizer, which Anthropic does not publish, so Claude is counted only by Anthropic's counting endpoint and never estimated from a GPT count.

Why do Claude models count differently from each other?

Anthropic introduced a newer tokenizer with Claude Opus 4.7, and by its own account the same text comes out at about 30 percent more tokens than on earlier models. That is why the count runs against the model you pick rather than one number for all of Claude.

What does an MCP tool list cost?

A client sends the definition of every tool on every request, whether the agent uses a tool or not. Paste the answer of tools/list and the page counts the whole list and each tool on its own, so you can see which descriptions are worth shortening. In our own measurement Playwright MCP's 24 tools cost 3,162 tokens per request.

What does one step of a browser agent cost?

Usually far more than the tool list, because of what each answer carries. Before an agent could act on one Wikipedia page, Playwright MCP returned 604,785 tokens and Scalebrowser 2,771, counted with cl100k\_base on 8 September 2026. [The full measurement](https://scalebrowser.net/blog/playwright-mcp-tokens) covers five browser tools and three sites.

Is my text stored?

No. It goes to our server, is counted and is gone. Do not paste secrets into any web form, this one included.

[The full check. Reads all five layers and shows where they contradict each other.](https://scalebrowser.net/check)

What this one page cannot tell you

What a model writes back. Output tokens depend on the answer, so only the input can be counted ahead of time. Images and PDFs are not counted here.