unitcostai

Tool Use Cost

What does one tool-using agent run cost?

How this tool works

This calculator treats one completed tool-use run as the unit of analysis. Each model/tool step is priced as one model request plus any weighted external tool, API, search, browser, execution, connector, retrieval, cache, and infrastructure cost that belongs to that step.

Use this when

  • You need a cost floor for runs that call tools, APIs, browsers, search, or connectors.
  • You want tool output tokens and paid external tool costs visible before setting credits or caps.
  • You need a per-run chargeback before modeling human approval or multi-agent fan-out.

How It Works

  1. Choose the model and set average model/tool steps per completed run.
  2. Enter prompt, tool output, response, retrieval, refresh, cache, and external tool/API overhead assumptions.
  3. Compare cost per tool-use run, monthly spend, break-even run price, margin, and sensitivity drivers.

Formulas

tool_docs_refresh_share_usd = tool_docs_refresh_monthly_usd / monthly_tool_use_runs

cost_per_tool_use_run_usd = generation + retrieval + reranking + refresh_share + vector_db + cache + external_tool_api_or_infra

optional_margin_pct = (run_price_or_credit_usd - cost_per_tool_use_run_usd) / run_price_or_credit_usd * 100

Assumptions and Units

  • Currency: USD
  • Unit of analysis: one completed tool-use run
  • A model/tool step is one model request that may include tool selection, tool output, retry, or final synthesis
  • External tool/API cost is a user-entered weighted average per model/tool step
  • Pricing source: daily pricing snapshot in repo, no runtime scraping

Worked Example

Start with three model/tool steps per run: plan, inspect one tool result, and synthesize the answer. If only one of those steps calls a paid external API, divide that API fee across the average steps. Use Agent Run Cost when the tool fee is not material, or Coding Agent Cost per Task when accepted tasks and rework are the denominator.

FAQs

Is this a live tool-fee scraper? No. Model pricing comes from the dated pricing snapshot; external tool/API fees stay explicit because provider tool fees are not one stable cross-provider schema.

How should I enter paid tool calls that happen only sometimes? Use a weighted-average external tool/API cost per model/tool step so the result stays tied to real run traces.

Related resources: AI Workflow Cost, Agent Run Cost, Break-even Price, Compare Model Costs, Tool Use Cost, Agent Run Cost, How Much Does an AI Agent Cost?, How To Price AI Agent Usage With Credits, Caps, and Margin, What Is an AI Agent?, RAG Cost Components Explained, How Many Tokens Per Request?, Coding Agent Cost per Task.

Pricing snapshot: 2026-07-09Provider: OpenAIModel: GPT-5.6 Luna

Decision Signal

Healthy

Current tool-use margin is 77.2%. Use this to compare modeled tool-use cost with a per-run price, credit, or chargeback.

Step 1 Provider and Model

Switch model assumptions using prices from the selected snapshot.

Step 2 Quick Mode

Use plain-language assumptions first. Open Advanced assumptions only if needed.
Starting pointApply the generic baseline or a more conservative downside scenario. If neither is selected, you are working from custom inputs.
Quick workflowStart with the closest workflow shape, then fine-tune the assumptions below.

A short run with a planner step, one tool result, and a final model response.

Optional Advanced assumptions

Tune retrieval, reranking, tool docs refresh, vector lookups, caching, and external tool overhead.
Show advanced inputs

Use weighted-average values when only some steps call paid external tools.

Scenario actions

Copy scenario URL

Paste into ChatGPT or Claude, or share with a teammate.

Save and track this scenario

Track pricing drift on this scenario and get an email if the latest result changes.

How tracking works

After you click Save and track, we carry this exact calculator state into the tracked-scenarios page so you can sign in and confirm the save.

We save your assumptions and the pricing snapshot used for this result.

When a newer pricing snapshot lands, we recompute the same scenario, show what changed, and email you if the latest result moved.

1 tracked scenario free, then $12/mo or $120/yr for up to 25 tracked scenarios.

Cost per tool-use run

$0.0411

Gross margin

77.2%

Estimated monthly tool-use cost

$205.62

Estimated monthly gross profit

$694.38

Top Cost Drivers

Most sensitive variables when each is moved up by 10%.
Model/tool steps / run10.0%
Output Tokens1.2%
Cache Hit Rate-0.9%

Totals

Summary metrics for monthly unit economics and margin.
Cost per model/tool step
$0.01371
Cost per tool-use run
$0.0411
Gross margin %
77.2%
Break-even price
$0.0411
Cost per model/tool step$0.01371
Cost per tool-use run$0.0411
Gross margin %77.2%
Break-even price$0.0411

Component Breakdown (USD/tool-use run)

Each cost component is computed independently and summed.

Largest cost block: Infra.

GenerationModel input/output token spend for requests.
$0.0087
RetrievalExtra model input spend from retrieved context chunks.
$0
RerankingReranker cost based on docs scored per request.
$0
Embeddings IngestionAmortized per-user share of the fixed monthly corpus embedding refresh cost.
$0
Vector DbVector database query cost across all requests.
$0
CacheSavings from cache hits. Negative means lower total cost.
$-0.0036
InfraNon-model infra overhead per request.
$0.036
GenerationModel input/output token spend for requests.$0.0087
RetrievalExtra model input spend from retrieved context chunks.$0
RerankingReranker cost based on docs scored per request.$0
Embeddings IngestionAmortized per-user share of the fixed monthly corpus embedding refresh cost.$0
Vector DbVector database query cost across all requests.$0
CacheSavings from cache hits. Negative means lower total cost.$-0.0036
InfraNon-model infra overhead per request.$0.036
Sensitivity RankingChange in total cost when one variable is increased by 10%.
VariableDelta cost %
Model/tool steps / runUser activity level per month.10.0%
Output TokensGenerated tokens per request.1.2%
Cache Hit RateFraction of requests served by cache.-0.9%
Input TokensPrompt-side tokens per request.0.7%
Retrieved ChunksRetrieved chunk count per request.0.0%
Tokens Per ChunkAverage chunk size in tokens.0.0%
Rerank DocsDocs reranked per request.0.0%
Vector Queries Per RequestVector query count per request.0.0%
Monthly tool-use runsActive-user estimate used to amortize fixed monthly embedding refresh.0.0%

Assumptions and Units

Explicit assumptions to keep outputs reproducible and auditable.
  • CurrencyUSD
  • Token unittoken
  • Pricing snapshot2026-07-09
  • Selected model rowOpenAI/GPT-5.6 Luna
  • Volume basisBusiness totals and fixed monthly terms use monthly tool-use runs as the denominator
  • Embedding refreshAmortized per tool-use run from the fixed monthly tool docs, schema, or index refresh term
  • Cache componentNegative value means cost savings

Recommended Next Step

Use these links to lower top cost drivers without guessing.

If tool-use cost is high, inspect model/tool step count, serialized tool output, retrieval depth, and external tool fees before changing packaging.

Choose the next agent economics slice

Agent Run CostCoding Agent Cost per TaskBreak-even PriceTool Use CostHow To Price AI Agent Usage With Credits, Caps, and Margin

Compare infra providers

View Infra Recommendations

Sources and Snapshot

Pricing comes from the current dated snapshot.

Active Pricing Row

Selected model

OpenAI / GPT-5.6 Luna

  • Input tokens$1 / 1M
  • Output tokens$6 / 1M

Shared retrieval defaults

  • Embedding input$0.02 / 1M
  • Rerank docs$1 / 1K

Continue Analysis

Switch tools

Read guides