Tool Calling Token Overhead Calculator

Tool calling token overhead calculator

Tool calling (function calling) is powerful, but it has a tax that does not show up in your prompt. Every tool definition is serialized into the request and billed as input — on every call, used or not. Then each tool result is fed back into the context and billed again. This calculator quantifies that hidden overhead per call and across a day of traffic.

How it works

The overhead has two parts. The schema cost is the sum of every tool definition’s tokens, included on each request. The result cost is the tokens a tool’s output adds back into the follow-up call. Together:

overhead/call = (num_tools × avg_schema_tokens) + tool_result_tokens
cost/call     = overhead/call / 1,000,000 × price
cost/day      = cost/call × calls_per_day

Note the schema cost is paid even when no tool fires — exposing a large catalog of tools to every request is the most common source of silent overspend.

How large does the overhead get?

A tool schema with a name, description, and five parameters can easily consume 80–150 tokens. If you expose ten tools on every request, that is 800–1,500 tokens of pure overhead before your actual prompt begins. At a frontier model price, that translates to a meaningful fraction of your total cost — especially if the agent rarely uses more than one or two tools per call.

Tool results amplify the cost further. When a tool returns a large JSON payload — a full API response, a long database row, or a retrieved document — that payload re-enters the context as input tokens on the follow-up call. A 2,000-token tool result that gets returned on 50% of calls adds 1,000 tokens of average overhead per call-pair. At high volume, this is significant.

Worked illustration

For example: an agent application that exposes 8 tools averaging 120 tokens each, receives tool results of about 500 tokens on 60% of calls, and runs 100,000 calls per day on a model costing $1.00 per million tokens.

• Schema overhead per call: 8 × 120 = 960 tokens
• Expected result overhead per call: 0.6 × 500 = 300 tokens
• Total overhead per call: 1,260 tokens
• Daily overhead cost: 100,000 × 1,260 / 1,000,000 × $1.00 = $126 per day purely from tool plumbing

That cost compounds: at 30 days, it is $3,780 just in tool overhead before counting the actual prompt, response, and business logic tokens.

Common optimizations

Limit tool exposure to the current step. In a multi-step agent, a routing layer can decide which subset of tools is relevant and only include those. If step 3 of a workflow never needs the “send_email” tool, do not send its schema on step 3 calls.

Trim schemas aggressively. Models generally do not need exhaustive parameter documentation to use a tool correctly. A one-sentence description per parameter is usually sufficient. Removing redundant or rarely needed optional parameters cuts schema tokens directly.

Summarize tool results. If a database query returns 50 rows but the model only needs to know whether any rows matched and the top result, return a summary. The raw 50-row payload is expensive input on every follow-up turn.

Tips and notes

• Scope tools to the step. Only pass the tools relevant to the current task; a router can decide which subset to expose.
• Compress schemas. Drop optional fields and verbose descriptions — the model rarely needs them and you pay for every word, every call.
• Cap result size. Truncate large API responses before returning them as tool results; the model usually needs a summary, not the full payload.