Tool Call Overhead Calculator
The hidden token tax of tool definitions — per provider, per tool, per call, verified.
System prompt and built-in tool token counts verified 2026-06-11against Anthropic's tool-use documentation. Schema tokens are your estimate — they depend on description verbosity.
Tool overhead alone costs ≈ $319.07/month at 2,000 calls/day on Claude Sonnet 4.6 input pricing ($3/M), before a single token of conversation.
- Tool-use system prompt (auto/none)
- 497 tok
- Custom tool schemas (5 × 250)
- 1,250 tok
- Built-in tools
- 0 tok
- Total overhead per call
- 1,747 tok
- Monthly overhead tokens (× 2,000/day × 30.44)
- 106.4M
Tip: tool definitions are static prefix content — ideal for prompt caching, which cuts their effective input cost by up to 90% on cache hits.
How it works
Every request you make with tools enabled carries an invisible token surcharge: a tool-use system prompt injected by the provider, plus the full serialized schema of every tool you attach, plus fixed costs for built-in tools. This calculator adds it all up — per call and per month in dollars — using token counts published by Anthropic and verified on the date shown in the tool.
The system prompt component varies more than most people expect. Opus 4.8 is the leanest at 290 tokens with tool_choice auto, rising to 410 when you force a tool call. Opus 4.7 pays 675 to 804 tokens for the same capability. Sonnet 4.6 and Opus 4.6 sit at 497/589, while Sonnet 4.5 and Haiku 4.5 are effectively identical at 496/588. On top of that, Anthropic's built-in bash tool adds a fixed 245 input tokens and the text editor tool roughly 700 — costs you pay whenever those tools are attached, not just when they are used.
The component that scales with your design is the schema tax. Each custom tool's name, description and JSON schema is serialized into the prompt on every call. Five tools at an average of 250 tokens each is 1,250 tokens per request — at 2,000 calls a day on a frontier model, that single design decision is worth real money monthly. The calculator multiplies your tool count by your average schema size, adds the model's system prompt overhead and any built-in tools, and prices the total at verified per-token rates.
Use the output to make two decisions. First, model choice: if your workload is tool-heavy with small payloads, the overhead difference between model generations is a real fraction of your bill. Second, schema diet: the calculator makes it obvious when description verbosity, not conversation content, is what you are paying for. Pair it with prompt caching — tool definitions are static prefix content, the single best candidate for cache reads at a tenth of the fresh-input price.
Frequently asked questions
Where does tool-use overhead come from?
Three places. First, enabling tools injects a special tool-use system prompt — on Anthropic models that ranges from 290 to 804 tokens depending on the model and tool_choice setting. Second, every tool definition you attach (name, description, JSON schema) is serialized into the prompt on every call. Third, built-in tools carry their own fixed costs: the bash tool adds 245 input tokens and the text editor tool adds about 700. None of this appears in your visible prompt, but all of it is billed.
Why does tool_choice change the overhead?
When tool_choice is auto or none, the model gets a shorter instruction block because it merely needs to know tools exist and may be used. When tool_choice is any or a specific tool, the system prompt grows — on Opus 4.8 from 290 to 410 tokens, on Opus 4.7 from 675 to 804 — because the model receives additional forcing instructions. If you do not need to force a tool call, leaving tool_choice on auto is a small but free saving on every single request.
How accurate are these numbers?
The per-model system prompt token counts come directly from Anthropic's published tool-use documentation and were verified on 2026-06-11 — the verification date is shown in the tool itself. Schema token counts are your own estimate since they depend entirely on how verbose your descriptions are; a typical small tool runs 100-200 tokens, a complex one with nested objects can exceed 500. The monthly dollar figure uses our verified pricing dataset at one call-equivalent of overhead per request.
How do I reduce tool overhead?
Trim tool descriptions to what the model needs for selection, not full documentation — the description is loaded on every call whether the tool is used or not. Remove tools the agent rarely calls; ten attached tools at 200 tokens each is 2,000 tokens per request before you say a word. Use prompt caching: tool definitions sit at the start of the prompt and are ideal cache content, cutting their effective cost by up to 90% on cache hits.
Does this apply to OpenAI and Google models too?
The structure does — every provider serializes your tool schemas into the context and adds framework instructions — but the exact token counts in this calculator are Anthropic's published figures, and other vendors do not document theirs with the same precision. As a rule of thumb, schema tokens dominate once you attach more than a handful of tools, and that part of the math transfers directly to any provider.
FORG tracks this automatically across every agent session — live cost attribution, budgets, and alerts.
Start tracking with FORG