MCP token cost calculator
Paste the tools your MCP server exposes. See how many tokens they add to every model request, which tools cost the most, and what that costs per month on each model, with and without prompt caching.
Added to every request
715 tokens
429 for 2 tools, give or take 5%, plus 286 for Anthropic's tool system prompt
Per month, no caching
$42.90
$1.43 per 1,000 requests
Per month, 90% cache hits
$7.29
$0.24 per 1,000 requests
Estimate calibrated against Anthropic's token counting endpoint on 2026-10-13: within 5% for a whole server, up to 17% off for one small tool. Prices from Anthropic pricing, checked 2026-10-21. Cache misses are billed at the cache write price.
Tokens per tool on Claude Sonnet 5.5
- convert_document 322
- list_supported_formats 107
Every model, 30,000 requests a month
| Model | Tokens / request | No caching | 90% cache hits |
|---|---|---|---|
| Claude Opus 5.5 | 715 | $85.80 | $14.59 |
| Claude Fable 5.1 | 429 | $128.70 | $18.98 |
| Claude Sonnet 5.5 | 715 | $42.90 | $7.29 |
| Claude Haiku 5.5 | 715 | $2.15 | $0.46 |
| Claude Sonnet 4.6 | 823 | $74.07 | $15.93 |
| GPT-6 Astra | 247 | $74.10 | $15.93 |
| GPT-6.1 Sol | 247 | $14.82 | $2.52 |
| GPT-6 Luna | 247 | $0.74 | $0.16 |
| GPT-5.6 Sol (promo price) | 247 | $29.64 | $6.37 |
| GPT-5.6 Terra | 247 | $14.82 | $3.19 |
| GPT-5.6 Luna | 247 | $1.48 | $0.32 |
| GPT-5.4 mini | 247 | $5.56 | $1.06 |
| Gemini 4 Argon (2026 price) | 247 | $14.82 | $2.15 |
| Gemini 4 Argon (from 2026-12-01) | 247 | $29.64 | $4.30 |
| Gemini 3.1 Pro (preview) | 247 | $14.82 | $2.82 |
| Gemini 3.8 Flash (2026 price) | 247 | $5.56 | $1.06 |
| Gemini 3.8 Flash (from 2027-01-01) | 247 | $11.12 | $2.11 |
| Gemini 3.5 Flash-Lite | 247 | $2.22 | $0.42 |
| Gemini 3.1 Flash-Lite | 247 | $1.85 | $0.35 |
This is only the cost of the definitions. Messages, tool results, and output are billed on top.
The three examples, worked out
Cost of the tool definitions alone at 30,000 model requests a month (about 1,000 a day), with no caching and with 90% cache hits. Token counts include Anthropic's published tool system prompt on Claude models.
Dawloom MCP server (2 tools, real)
The live tools/list from dawloom.com/mcp, fetched 2026-10-13.
| Model | Tokens / request | No caching / month | 90% cache hits / month |
|---|---|---|---|
| Claude Opus 5.5 | 715 | $85.80 | $14.59 |
| Claude Sonnet 5.5 | 715 | $42.90 | $7.29 |
| Claude Haiku 5.5 | 715 | $2.15 | $0.46 |
| GPT-6.1 Sol | 247 | $14.82 | $2.52 |
| GPT-5.4 mini | 247 | $5.56 | $1.06 |
| Gemini 3.8 Flash (2026 price) | 247 | $5.56 | $1.06 |
Shop orders, short descriptions (6 tools)
An example we wrote. Six order tools with one-line descriptions.
| Model | Tokens / request | No caching / month | 90% cache hits / month |
|---|---|---|---|
| Claude Opus 5.5 | 1,076 | $129.12 | $21.95 |
| Claude Sonnet 5.5 | 1,076 | $64.56 | $10.98 |
| Claude Haiku 5.5 | 1,076 | $3.23 | $0.69 |
| GPT-6.1 Sol | 366 | $21.96 | $3.73 |
| GPT-5.4 mini | 366 | $8.24 | $1.56 |
| Gemini 3.8 Flash (2026 price) | 366 | $8.24 | $1.56 |
Shop orders, long descriptions (6 tools)
The same six tools with long descriptions, enums, and patterns.
| Model | Tokens / request | No caching / month | 90% cache hits / month |
|---|---|---|---|
| Claude Opus 5.5 | 2,513 | $301.56 | $51.27 |
| Claude Sonnet 5.5 | 2,513 | $150.78 | $25.63 |
| Claude Haiku 5.5 | 2,513 | $7.54 | $1.62 |
| GPT-6.1 Sol | 1,338 | $80.28 | $13.65 |
| GPT-5.4 mini | 1,338 | $30.11 | $5.72 |
| Gemini 3.8 Flash (2026 price) | 1,338 | $30.11 | $5.72 |
Prices from Anthropic, OpenAI, and Google, checked 2026-10-21.
How the estimate works
No tokenizer ships with this page. Each tool is cut down to its name, description, and input schema, then minified. The estimate is a fixed cost per tool, plus a weight for each JSON punctuation mark, plus a weight for every other character.
We fit the weights on 2026-10-13 against Anthropic's token counting endpoint (Claude Sonnet 5.5 and Sonnet 4.6) and OpenAI's open o200k tokenizer, over 20 tool definitions. Plain "characters divided by 4" was 42% low on average for the newer Claude tokenizer. Our fit was within 3.5% on every example server.
What it doesn't count
Only the tool definitions. Tool calls, tool results, the conversation, and output are billed on top, and tool results are often bigger than the definitions. Some clients also add their own instructions per server.
The full method, the calibration numbers, and how to cut the overhead are in our guide, why your MCP server uses so many tokens.
Questions clients ask
Short answers, no sales language. If yours isn't here, ask an engineer directly.
Ask us anythingWhy is my MCP server using so many tokens?
Because the client sends every tool's name, description, and input schema to the model on every request, even when no tool gets called. Long descriptions, big enums, and many tools add up. Anthropic's tool search docs give a five-server example (GitHub, Slack, Sentry, Grafana, Splunk) that uses about 55,000 tokens before Claude does any work.
How accurate is the estimate?
For Claude models, we checked it against Anthropic's token counting endpoint on 2026-10-13 across 20 tool definitions. Whole example servers came within 3.5% of the real count. A single small tool can be off by up to 17%. For GPT and Gemini, the providers don't publish how tool definitions are counted, so treat the figure as give or take 25%.
Does prompt caching fix it?
It cuts the price, not the tokens. Tool definitions sit at the very start of the prompt, so they cache well if the list stays identical between requests. A cache hit costs 10% or less of the input price on most current models. The tokens still fill the context window.
How do I paste my server’s tools?
Call tools/list on your server and paste the JSON response, or paste the tools array from your code. You can try it on our own server: POST {"jsonrpc":"2.0","id":1,"method":"tools/list"} to https://dawloom.com/mcp.
Need an MCP server that stays lean?
We build MCP servers with a small, well-described tool list. An engineer replies, not a sales team.