MCP server token cost: what tool definitions add
Short answer: MCP server burns about 35K input tokens (55% cached) and 2,500 output tokens per typical task, so one task costs $0.0561 on Grok 4.7. At 400 tasks a month that is $22.45; without caching the same task would cost $0.0850. Evidence: measured.
Cost per task, three price points
One row per scenario from our tool data. The three price columns price the identical token counts on the cheapest, mid and premium reference models, so the difference is the model rate, not the workload.
| Scenario | Input tokens | Cached input | Output tokens | Cheapest current model | Mid-tier reference model | Premium reference model |
|---|---|---|---|---|---|---|
| Light task one short request with a small context and one or two tool calls | 8K | 40% | 800 | $0.000912 | $0.0160 | $0.1824 |
| Typical task one coding or agent task with several tool calls and a medium context | 35K | 55% | 2,500 | $0.00302 | $0.0561 | $0.6035 |
| Heavy session a multi-turn agent session with a large repo or long history; input counts all tokens re-sent across turns | 150K | 65% | 8K | $0.0102 | $0.2018 | $2.04 |
Rates: input $2 / cached $0.5 / output $6 per 1M on Grok 4.7. Prices checked 2026-11-18.
Monthly budget
| Scenario at 400 tasks / month | Cheapest current model | Mid-tier reference model | Premium reference model |
|---|---|---|---|
| Light task | $0.3648 | $6.40 | $72.96 |
| Typical task | $1.21 | $22.45 | $241 |
| Heavy session | $4.09 | $80.70 | $818 |
At 400 tasks a month, one task at the typical size costs $0.0561 on Grok 4.7, so $100 covers about 1782 tasks. The same budget buys 166 tasks on the premium reference and 33140 on the cheapest one.
Server-rendered from the typical task on Grok 4.7 at 400 tasks a month. Change a field to recalculate; nothing is sent anywhere and the formula is the one under Method.
Where the tokens go
- 35K input tokens, 55% of them cached. Every connected MCP server contributes tool definitions and descriptions to the context before the model does anything, and each tool call's result is fed back into the next request. Combining several servers multiplies the fixed overhead, which is why client-side tool search or on-demand tool loading matters. of the input is the tool's prompt, tool schema and history; the cached share is what the provider can reuse on the next step without re-reading it at the full rate.
- 2,500 output tokens — 14.0× fewer tokens than the input. The answer is short relative to the context the tool sends with it, which is the normal shape for an agent step.
- Output is 7% of the token count but 27% of the bill on Grok 4.7. Output is priced higher than input, so the visible answer still matters even when most tokens are input.
- Per 1,000 tasks: $56.12. That is the number to put in a budget line, not the per-token rate: 1,000 tasks at the typical size cost $56.12 on Grok 4.7 and $604 on GPT-6 Cyber.
- At 400 tasks a month: $22.45 on Grok 4.7 and $241 on GPT-6 Cyber. The premium reference is 10.8× the mid-tier bill; the cheapest reference is $1.21.
The provider basis behind these numbers
Anthropic's engineering write-up measured the problem: MCP clients load tool definitions upfront, agents connected to thousands of tools need to process hundreds of thousands of tokens before reading a request, and every intermediate tool result passes back through context (a two-hour meeting transcript can add ~50k tokens). Their code-execution approach cut a representative case from 150,000 to 2,000 tokens (98.7%). We model a typical single server with 15-25 tools at roughly 100-400 tokens of schema per tool plus tool results; light/task/session tiers scale the number of connected servers and tool calls.
What caching is worth here
Our data does not assume a cached prefix for this tool (55% cached share), so the task price above is already the uncached input price. If the tool does repeat a stable prefix, adding it to the provider cache cuts the input side to $0.5 per 1M on Grok 4.7.
How to reduce it
- Connect only the servers a task needs and disable the rest.
- Prefer clients with deferred tool loading or tool search so schemas load on demand.
- Make tool descriptions short and avoid returning large raw payloads; filter results server-side.
- Prefer a CLI tool over an MCP server when one already exists, since CLI tools add no per-tool listing.
Each tip changes a different part of the formula: caching changes the input rate, trimming history changes the input count, and routing simple steps to a cheaper model changes the rate for part of the work.
Sources and evidence
| Source | Read on | What it establishes |
|---|---|---|
| www.anthropic.com | 2026-11-04 | Anthropic engineering analysis with a measured 150,000 -> 2,000 token context reduction for tool-heavy MCP workflows. |
| modelcontextprotocol.io | 2026-11-04 | MCP client best practices covering tool discovery, context management and pagination. |
Evidence label: Measured
Anthropic's engineering write-up measured the problem: MCP clients load tool definitions upfront, agents connected to thousands of tools need to process hundreds of thousands of tokens before reading a request, and every intermediate tool result passes back through context (a two-hour meeting transcript can add ~50k tokens). Their code-execution approach cut a representative case from 150,000 to 2,000 tokens (98.7%). We model a typical single server with 15-25 tools at roughly 100-400 tokens of schema per tool plus tool results; light/task/session tiers scale the number of connected servers and tool calls.
Method
The formula
Cost per task = (uncached input × input rate + cached input × cached rate + output × output rate) ÷ 1,000,000. Rates are the standard list prices in the price table (checked 2026-11-18); token counts come from src/tools.json (tool data reviewed 2026-11-05).
Why three reference models
The columns are the cheapest, median and most expensive current models by the cost of one typical agent task, named above so you can check each rate. They are a range, not a prediction of which model MCP server uses.
What the monthly figure assumes
400 tasks a month (the tasks-per-month field in our tool data), the same count for every scenario, at the standard rate with no batch discount. A heavy session can count as several tasks, so replace the number with your own before you put it in a budget.
Official links
Related skills & plugins
Other entries in the MCP servers category. Each name links to its token-cost page or its official site.
- MCP Registry — Official registry for discovering and publishing MCP servers. (Resource)
- Claude Code plugins — Claude Code's installable extension format beyond skills or MCP servers. (Resource)
- Claude Skills — Claude Code skills that load only when invoked and cap at 5k tokens. ($0.0444 per task, Documented)
- All skills, plugins and MCP token costs
Related models
The three reference models in the cost table, with their full price tiers.
- GPT-6 Luna — $0.1 input / $0.5 output per 1M tokens, $0.00302 per typical task in the table above
- Grok 4.7 — $2 input / $6 output per 1M tokens, $0.0561 per typical task in the table above
- GPT-6 Cyber — $20 input / $100 output per 1M tokens, $0.6035 per typical task in the table above
More
- Benchmark comparison: price vs performance
- Best LLM for coding in 2026
- Find a model for your use case
- Compare two to four models on price
Related
- All skills, plugins and MCP token costs
- Best LLM for coding in 2026
- Find a model for your use case
- Token cost calculator
- Compare two models on price
- Claude Code token cost
- OpenAI Codex CLI token cost
Questions
How much does MCP server cost per task?
One typical task costs $0.0561 on Grok 4.7, $0.00302 on the cheapest reference model (GPT-6 Luna) and $0.6035 on the premium reference (GPT-6 Cyber). The spread is the model rate: the token counts are identical in all three columns.
What does MCP server cost per month at 400 tasks?
$22.45 a month on Grok 4.7 at the typical task size. Without the 55% cached input assumption it would be $34.00. Per 1,000 tasks the token bill is $56.12.
Why does MCP server use so many input tokens?
Every connected MCP server contributes tool definitions and descriptions to the context before the model does anything, and each tool call's result is fed back into the next request. Combining several servers multiplies the fixed overhead, which is why client-side tool search or on-demand tool loading matters. The token shape is 35K input tokens (55% cached) against 2,500 output per typical task; input is resent every step, output is not.
Are the MCP server token numbers measured or estimated?
The evidence label is "Measured". Anthropic's engineering write-up measured the problem: MCP clients load tool definitions upfront, agents connected to thousands of tools need to process hundreds of thousands of tokens before reading a request, and every intermediate tool result passes back through context (a two-hour meeting transcript can add ~50k tokens). Their code-execution approach cut a representative case from 150,000 to 2,000 tokens (98.7%). We model a typical single server with 15-25 tools at roughly 100-400 tokens of schema per tool plus tool results; light/task/session tiers scale the number of connected servers and tool calls. Sources and read dates are in the table above; if a number is an estimate, the basis says what it assumes, and it is never described as a measurement.
Next step → See every tool cost side by side
Price data last checked 2026-11-18 against official sources · how we verify
Price-change alerts
One short email when a provider changes a price or a promo ends. No spam, unsubscribe any time.