What does every agent task cost?
Arcane Cost Center attributes token use and spend to the agent, task, model, team, and authority behind it. It lets organizations track AI costs and optimize execution with task outcome and quality in view.
How can teams inspect AI token use and spend?
Cost Center presents attributed usage as an operational view, not a provider invoice. Teams can change the time range, compare spend with token volume, group execution by task, agent, or model, and inspect the runs behind a cost change.
Usage over time
Provider usage joined to Arcane task records
Spend by provider
30-day attributed cost
Cost attribution explorer
Select a row to inspect its execution profile
Most variance comes from retries after tool timeouts in finance-agent@v1.6.
Higher cache reuse could reduce research brief input cost without changing the selected model.
What is AI agent cost tracking?
AI agent cost tracking connects model usage and execution spend to the work that caused it. Arcane attaches that data to the same task and decision context used for identity, permissions, and audit.
- Usage
- Input · output · cached tokens
- Execution
- Calls · retries · task steps
- Attribution
- Agent · task · model · team
- Evaluation
- Cost · outcome · quality
How does Cost Center make AI spend actionable?
It turns raw usage into an attributable task record, then makes comparable dimensions available for analysis, budgets, and anomaly review. Provider-reported fields are preserved when available rather than estimated as if they were observed facts.
Know what created the spend
Tie provider usage and cost to the exact task, agent, team, model, and delegating authority instead of leaving it in an aggregate cloud bill.
Measure the whole task
Roll up model calls, retries, and tool-driven execution under one task so a multi-step agent run has one attributable cost record.
Find expensive patterns
Compare agents, tasks, models, and teams to identify high-token workflows, repeated retries, and cost changes over time.
Optimize for value, not tokens alone
Evaluate cost beside task outcome and quality signals so lower spend does not quietly produce less accurate or less useful work.
Which cost dimensions can teams analyze?
Cost Center provides organization-wide totals without discarding the agent and task context needed to explain them.
| Dimension | What it explains |
|---|---|
| Organization | Total AI execution cost across teams and environments. |
| Team or cost center | Spend attributed to the group responsible for the work. |
| Agent | Token use and cost for a registered agent and its versions. |
| Task | The complete cost of pursuing a defined objective across model and tool calls. |
| Model | Usage by provider, model, and model configuration. |
| Authority | Cost connected to the user or system that delegated the task. |
| Outcome | Cost evaluated beside completion, quality, and accuracy signals. |
How is the cost of an AI agent task calculated?
A task cost is the sum of attributable execution charges across every model call, retry, and metered tool step in that task. Provider-reported cost is preserved when available; calculated cost must retain the usage source, pricing version, currency, and calculation time.
task cost = Σ model-call charges + Σ recorded tool chargesEach model-call charge separates input, cached-input, and output usage according to the applicable provider price. Retries remain part of the task total because they consumed resources while pursuing the same objective.
- Usage source
- Provider response or observed meter
- Pricing input
- Provider, model, region, and effective version
- Task binding
- Stable task and execution identifiers
- Result
- Currency, reported or calculated cost, and timestamp
| Execution | Tokens | Cost |
|---|---|---|
| Primary model call | 38,420 | $0.84 |
| Retry | 6,180 | $0.13 |
| Recorded tool charge | — | $0.04 |
| Task total | 44,600 | $1.01 |
How should an organization optimize token use?
Start with the task, not the invoice. Compare the cost of equivalent tasks, inspect unusually expensive runs, and evaluate any model or prompt change against completion and quality signals before adopting it.
Measure
Capture usage and cost under one task identifier.
Compare
Group equivalent work by agent, model, team, and outcome.
Improve
Reduce waste while checking that task quality still holds.