GPT-5.5 Tool Schema Cost
Piling tools onto GPT-5.5? Every call re-sends all their schemas as input tokens. This page computes exactly how much of your bill that is — and how far caching cuts it.
For engineers piling tools onto an agent — computes how much of every bill is the tool/function schemas you re-send on each call, and how far prompt caching cuts it, not a 'tools are free' assumption.
Model prices from OpenRouter · updated 2026-07-13
01Your tools & workload
$5/M in · $30/M out · cache reads $0.5/M
02 The hidden per-call tax
Your 20 tools add $900/month (37% of input) — small now, but it grows with every tool.
$0.01 prompt+output + $0.001800 tool schema = $0.02/call
Without caching the schema tax would be $0.02/call — caching the stable schemas is your biggest lever.
Related cost calculators
Prices from OpenRouter, snapshot 2026-07-13, synced daily. Per call = prompt × in + (numTools × schemaTokens) × [cache-read when caching, else input] + output × out, per 1M tokens. Providers count tool-schema tokens slightly differently and a few trim unused tools — treat this as the standard “all tools, every call” upper bound. All math runs in your browser.
How the math works
Defining 20 tools on GPT-5.5 looks free — you write the schema once. But every API call re-sends all 20 tool definitions (names, descriptions, parameter JSON schemas) as input tokens: 180 tokens each here, 3,600 tokens of fixed overhead bolted onto each request, on every call, forever.
Same baseline, one identity: the naive per-call cost (your 600-token prompt + output) is $0.01. The tool schemas add $0.001800 on top — the real per-call cost is $0.02. At 500,000 calls/month that hidden tax is $900/month you'd never see on a per-call price sheet.
The tax scales with how many tools you've defined, not how many a given call uses — a kitchen-sink agent pays for all of them on every request. Here the tool schemas are 37% of your input tokens; keep an eye on it as the tool catalog grows — the share climbs with every tool added.
Prompt caching is the lever, because tool schemas are stable between calls — the ideal cacheable prefix. With caching on, the schemas re-read at $0.5/M instead of $5/M, cutting the tax to $0.001800/call. Turn caching off and it jumps to $0.02/call (133% on top of your real prompt).
Not modeled: providers count tool-schema tokens slightly differently than raw text, and a few trim unused tools automatically — treat this as the upper bound of the standard "all tools, every call" behavior. Inference prices sync daily from OpenRouter (updated 2026-07-13); this is a token-accounting comparison, not a separate price source. All math runs client-side with tested code.
Frequently asked questions
How much do a GPT-5.5 agent's tool definitions actually cost?
At this page's defaults — 20 tools, 180 tokens each, 500,000 calls/month — the re-sent schemas cost $900/month (13% on top of your actual prompt-and-output cost). Tune the sliders for your real tool catalog.
Why do tools cost anything if I only call one per request?
Because the model has to be told about every tool it could call — all 20 schemas are sent as input on every request, whether the call uses one tool or none. That's 3,600 tokens of fixed overhead per call here, independent of which tool actually fires.
Does prompt caching fix this?
Mostly — tool schemas are identical between calls, so they're the perfect cacheable prefix. Caching drops the per-call tax from $0.02 (full price) to $0.001800 (cache-read) here. It doesn't eliminate the tax, but it shrinks it a lot.
What's the cheapest way to cut this?
Two levers: trim tools the agent rarely uses (the tax is linear in tool count), and enable prompt caching so the stable schemas read at the discounted rate. Splitting a giant tool catalog into narrower agents that each carry fewer tools also helps — each call then pays for a smaller schema.
Are these prices current?
Inference prices sync daily from OpenRouter (updated 2026-07-13). This mold adds no separate hand-verified price source — it's a token-accounting comparison (prompt vs re-sent tool schemas) on top of the live catalog, so it stays accurate as prices change automatically.
Should I worry about this for a GPT-5.5 agent?
It's a smaller slice here (37% of input), but it grows with every tool you add and every call you make — worth watching as the agent's tool catalog expands.